Last 7 Days (July 13 – July 19, 2026)
Robotic manipulation is inherently multi-frame: local actions may be simple in an end-effector frame, while transport, upright-object handling, and whole-body coordination are better represented in a base-aligned frame. However, modern diffusion-based visuomotor policies typically commit to a single predefined action frame, forcing one denoiser to model action distributions that are often unnecessarily complex in that frame. We propose Mixture of Frames Policy (MoF), a diffusion policy that performs synchronized action denoising across multiple coordinate frames. MoF maintains a single canonical diffusion state, re-expresses it in several task-relevant frames, applies frame-specialized denoisers, and fuses their noise predictions back in the canonical frame. To make this possible for intermediate noisy diffusion states, we introduce a column-based 6D rotation representation within an SE(3) action parameterization that supports exact, differentiable frame transformations without requiring noisy rotations to lie on the SO(3) manifold. Across nine simulated bimanual manipulation tasks, we show that the best action frame is task-dependent and that MoF improves over oracle frame selection and standard Mixture-of-Experts (MoE) baselines. We further evaluate MoF on two real-world bimanual mobile manipulation tasks, demonstrating that it outperforms all constituent single-frame baselines. Project homepage: https://mofpo.github.io
Primary: Stanford University
All Institutions: Stanford University, Apple, Toyota Research Institute
The paper presents a highly innovative and technically sound approach to improving diffusion-based visuomotor policies by leveraging multi-frame reasoning. The introduction of a column-based 6D rotation representation for exact noisy state transformation is a significant methodological advance. Comprehensive experiments in simulation and on real-world robots demonstrate substantial improvements over strong baselines, including oracle frame selection. The work is well-motivated, rigorously evaluated, and addresses a fundamental limitation in current robotic policy architectures. It is a strong candidate for top-tier publication and will likely influence future research in robotic policy learning.
The paper proposes a novel architecture for diffusion-based visuomotor policies called Mixture of Frames (MoF). The core insight is that robotic manipulation actions are naturally represented in different coordinate frames at different stages of a task (e.g., end-effector for grasping, base for transport). Standard diffusion policies fix a single action frame, forcing the denoiser to model complex, frame-dependent distributions. MoF maintains a single canonical diffusion state but applies frame-specialized denoisers in their respective frames, fusing the noise predictions back into the canonical frame. A critical technical contribution is the introduction of a column-based 6D rotation representation for SE(3) actions. This allows for exact, differentiable frame transformations of *noisy* intermediate states without requiring projection onto the SO(3) manifold, which is non-linear and lossy. This mathematical insight is elegant and solves a significant implementation hurdle for multi-frame reasoning in diffusion policies. The method is generalizable to any diffusion policy and can be implemented as a drop-in replacement or extension.
The evaluation is comprehensive and rigorous. The authors first establish the motivation by showing that action frame choice has a substantial, task-dependent impact on performance in bimanual mobile manipulation, with a 15% gap between the best and worst single frames. They then demonstrate that MoF outperforms oracle frame selection (which knows the best frame per task) and standard Mixture-of-Experts (MoE) baselines that do not operate in multiple action frames. The ablation studies are particularly strong, showing that the proposed column-based representation is essential (orthogonalization leads to collapse in ensemble variants) and that the router effectively switches between frames based on task phase. Real-world experiments on two complex bimanual mobile manipulation tasks (Pouring, Serving) further validate the approach, showing significant improvements over single-frame baselines. The use of simulated benchmarks (BiGym, DexMimicGen) and real-world validation provides a robust evidence base.
The paper provides detailed descriptions of the method, including the specific rotation representation and the fusion mechanism. The project homepage link suggests code availability. The experimental setup is well-described, including dataset sizes, training epochs, and evaluation metrics. The ablation studies provide clear evidence of component contributions. The reliance on standard simulators (BiGym) and real-world hardware (HoMMI) enhances reproducibility.
The method relies on a designer-specified set of candidate frames. While the router learns to weight them, the set must be predefined. The paper acknowledges this and suggests future work on learning or discovering frames. Additionally, the router is supervised via denoising loss, which may not perfectly align with policy performance, though the results suggest it works well in practice. The current experiments focus on smaller diffusion policies; scaling to large VLA models is noted as future work.
This work has significant implications for robotic manipulation, particularly in complex, multi-stage tasks where different coordinate systems are advantageous. By improving the robustness and performance of visuomotor policies, it contributes to the broader goal of autonomous robots capable of operating in unstructured environments. The technical contribution regarding SE(3) transformations is also valuable for the broader machine learning community working with geometric deep learning and diffusion models. The paper presents a highly innovative and technically sound approach to improving diffusion-based visuomotor policies by leveraging multi-frame reasoning. The introduction of a column-based 6D rotation representation for exact noisy state transformation is a significant methodological advance. Comprehensive experiments in simulation and on real-world robots demonstrate substantial improvements over strong baselines, including oracle frame selection. The work is well-motivated, rigorously evaluated, and addresses a fundamental limitation in current robotic policy architectures. It is a strong candidate for top-tier publication and will likely influence future research in robotic policy learning.
Compression is fundamental to intelligence. A model that can represent its training data as a short code has discovered regularities that enable generalization. Large neural networks may learn functions far simpler than their parameter counts suggest, but it is challenging to construct codes that realize this simplicity. Parameter-based methods such as quantization produce code lengths that scale with model size, insensitive to how much information the parameters store. Prequential coding bypasses this issue by compressing the training trajectory, but codes the exact data sequence regardless of how much the model learns, yielding large codes when the data has high entropy. We introduce requential coding, where a teacher model selects training samples drawn from the student's own distribution. The student's code records only these selections, which cost bits only where teacher and student disagree. The resulting code length is independent of parameter count and data entropy, and often orders of magnitude shorter than the prequential counterpart, with an advantage that grows with scale. This compression sheds light on phenomena inaccessible to prior compressors. Holding loss fixed, larger models and ensembles compress to much smaller sizes despite more parameters. Plugged into a PAC-Bayes bound, the requential code yields state-of-the-art generalization guarantees for billion-parameter LLMs, outperforming bounds built on aggressive post-training quantization even granted zero error. The bound tightens with scale in the compute-optimal regime, as models become increasingly compressible relative to dataset size. The same code predicts that models gradually overfit when trained for multiple epochs. It also isolates the learnable information in a dataset from its unpredictable, random content, revealing that lower-entropy text holds far more learnable structure than higher-entropy image data.
Primary: New York University
All Institutions: Carnegie Mellon University, New York University
Requential coding has a profound broader impact on the field of machine learning: 1. **Understanding Generalization**: It provides a rigorous, operational measure of model complexity that directly correlates with generalization performance, especially for large neural networks where traditional measures (like parameter count) are misleading. This is a critical step towards a deeper theoretical understanding of deep learning. 2. **Scaling Laws**: The finding that larger models are more compressible at fixed loss offers crucial insights into the efficiency of scaling, suggesting that larger models indeed learn simpler, more generalizable functions. This can guide future research into model architectures and training paradigms. 3. **PAC-Bayes Bounds**: Achieving state-of-the-art, non-vacuous PAC-Bayes generalization bounds for billion-parameter LLMs is a major theoretical breakthrough, making these bounds relevant for modern large-scale models. 4. **Data Curation and Selection**: The ability to distinguish learnable information from random content in datasets provides a principled metric for data quality and could inform strategies for more effective data curation and selection. 5. **Theoretical Tool**: Requential coding serves as a powerful analytical tool for diagnosing phenomena like overfitting and understanding the information flow during training, opening new avenues for theoretical and empirical investigation in deep learning. Requential coding introduces a novel, highly efficient model compression scheme that significantly advances the understanding of generalization in deep learning by providing a rigorous measure of model complexity independent of parameter count and data entropy. This method yields state-of-the-art PAC-Bayes generalization bounds for billion-parameter LLMs, reveals that larger models are more compressible, and offers new insights into overfitting and the learnable information content of datasets, thus serving as a powerful analytical tool for the field.
The paper introduces "requential coding," a novel and highly efficient method for compressing generative models. It builds upon the principles of prequential coding and relative entropy coding (REC) but addresses their fundamental limitations. Unlike parameter-based methods (e.g., quantization) that scale with model size regardless of information stored, or prequential coding that scales with dataset size and entropy, requential coding decouples code length from both. The core idea is that a "student" model generates candidate training samples from its own distribution, and a stronger "teacher" model selects one of these candidates. The student's code then records only the index of the selected sample, which costs approximately $KL(Q_t\|P_t)$ bits, where $Q_t$ is the teacher and $P_t$ is the student. This effectively distills the teacher's knowledge into the student by efficiently transmitting only the "disagreement" between them. The methodology is rigorously defined, including the encoder and decoder protocols, and a formal bound for the expected code length is derived in the appendix, showing it scales with the cumulative teacher-student KL divergence. Practical improvements like "teacher smoothing" (using an EMA of teacher checkpoints) and "iso-loss projection" (periodically resetting the teacher to the student's performance level) are introduced to further reduce the code length without sacrificing performance. The paper also provides a detailed analysis of the runtime for code length evaluation, encoding, and decoding, acknowledging the computational cost of actual transmission (which can be high due to REC's proposal generation) but emphasizing that evaluation is much more efficient. The use of a universal integer code (Elias delta) for the index ensures the decoder doesn't need prior knowledge of the KL divergence. This is a sophisticated and well-thought-out approach that significantly advances the state of the art in model compression theory.
The experimental evaluation is comprehensive and compelling, demonstrating the power of requential coding across various tasks and scales. 1. **Compression Benchmarking**: Requential coding is benchmarked against prequential coding and idealized 4-bit post-training quantization (PTQ) on autoregressive transformers (100M parameters) trained on OpenWebText, CIFAR-5M, and FineWeb. It achieves one to two orders of magnitude shorter codes than prequential coding, dominating the Pareto frontier of loss vs. code length. The prequential heuristic (a non-rigorous estimate) also falls well above requential coding. 2. **Scaling Insights**: * **Model Size**: Experiments show that larger models (from 10M to 1B parameters) are *more compressible* at a fixed loss, despite having more parameters. This is a crucial finding, as it suggests larger models learn simpler functions relative to their capacity, which parameter-based methods fail to capture. * **Ensembles**: Larger ensembles are also shown to be more compressible, reaching lower loss at a fixed code budget. 3. **Generalization Guarantees (PAC-Bayes Bounds)**: This is a highlight. Plugging the requential code length into a PAC-Bayes bound yields state-of-the-art generalization guarantees for billion-parameter LLMs. These bounds: * Improve with model scale, mirroring the true test loss, for models trained on a fixed amount of data. * Outperform even *idealized lossless 4-bit PTQ bounds* in the compute-optimal scaling regime ($D=20N$), a significant breakthrough as previous non-vacuous bounds for LLMs were based on realistic PTQ. * Tighten with scale, as the code length per parameter decays as a power law, predicting a vanishing generalization gap. 4. **Overfitting Prediction**: The requential code successfully predicts gradual overfitting under data repetition, showing the divergence between training and test loss as the complexity penalty rises. 5. **Learnable Information Content**: The method effectively isolates learnable structure from random information in datasets, revealing that lower-entropy text holds more learnable structure than higher-entropy image data, and distinguishing it from random or trivially repeating strings. This provides a principled way to rank datasets by their "epiplexity." The experiments are well-designed, the results are clearly presented (e.g., Figure 3, 4, 5), and the conclusions drawn are profound, offering new insights into deep learning phenomena that were previously inaccessible to rigorous measurement.
The paper explicitly states that "Code available at https://github.com/shikaiqiu/requential-coding". The methodology is described with sufficient detail, including pseudocode for REC and descriptions of teacher smoothing and iso-loss projection. The appendix provides full experiment details, including how code lengths are computed, datasets used, and per-figure configurations. This commitment to open-sourcing the code and providing detailed experimental setups significantly enhances reproducibility.
The authors acknowledge several limitations: 1. **Lossy Compression of Teacher**: Requential coding provides a lossless compression for the student model, which is a distillation of the teacher. If the goal is to compress a pre-trained model not trained via distillation, it becomes a lossy compression of that original model. 2. **Encoding Runtime**: While code length *evaluation* is efficient, *actually transmitting* a model via requential coding can be prohibitively slow due to the exponential scaling of REC encoding time with KL divergence. The paper primarily presents it as a tool for *evaluation* and *understanding* rather than practical transmission. 3. **Information Forgetting**: The current code length only grows with training steps and doesn't account for information that might be learned early but subsequently forgotten. Leveraging forgotten information could potentially further reduce the code length. 4. **Teacher Optimization**: The choice of teacher sequence is simple, and further gains are expected from optimizing it.
Requential coding has a profound broader impact on the field of machine learning: 1. **Understanding Generalization**: It provides a rigorous, operational measure of model complexity that directly correlates with generalization performance, especially for large neural networks where traditional measures (like parameter count) are misleading. This is a critical step towards a deeper theoretical understanding of deep learning. 2. **Scaling Laws**: The finding that larger models are more compressible at fixed loss offers crucial insights into the efficiency of scaling, suggesting that larger models indeed learn simpler, more generalizable functions. This can guide future research into model architectures and training paradigms. 3. **PAC-Bayes Bounds**: Achieving state-of-the-art, non-vacuous PAC-Bayes generalization bounds for billion-parameter LLMs is a major theoretical breakthrough, making these bounds relevant for modern large-scale models. 4. **Data Curation and Selection**: The ability to distinguish learnable information from random content in datasets provides a principled metric for data quality and could inform strategies for more effective data curation and selection. 5. **Theoretical Tool**: Requential coding serves as a powerful analytical tool for diagnosing phenomena like overfitting and understanding the information flow during training, opening new avenues for theoretical and empirical investigation in deep learning. Requential coding introduces a novel, highly efficient model compression scheme that significantly advances the understanding of generalization in deep learning by providing a rigorous measure of model complexity independent of parameter count and data entropy. This method yields state-of-the-art PAC-Bayes generalization bounds for billion-parameter LLMs, reveals that larger models are more compressible, and offers new insights into overfitting and the learnable information content of datasets, thus serving as a powerful analytical tool for the field.
Mainstream visual encoders are pretrained on natural images and cannot be effectively applied to document images without document-oriented adaptation, as dense text and fine-grained character strokes demand character-level visual perception. We present MonkeyOCRv2, a visual-text pretrained model for document AI. First, we construct MonkeyDoc v2, to our knowledge the largest document-image pretraining corpus, comprising 113 million images spanning 17 languages. Second, we propose a pretraining strategy that jointly learns image-to-text generation and pixel-level document reconstruction: the former aligns visual representations with textual content, while the latter preserves character strokes and layout details. Extensive experiments are conducted on five representative document analysis tasks, including text recognition, formula recognition, text detection, document tampering detection, and overlapping text segmentation. Replacing the original encoders with MonkeyOCRv2 consistently improves performance across all five tasks. Finally, we validate its effectiveness as the vision encoder of multimodal large language models on the more challenging tasks of document parsing and document understanding. Kept frozen and paired with a lightweight language model, it yields a 0.7B document parsing model that sets a new open-source state-of-the-art on MDPBench, a recent benchmark spanning digital-born and photographed documents across 17 languages, surpassing the previous best 3B dots.mocr by 2.8% absolute with a vision encoder roughly 11times smaller. The frozen encoder also powers a document understanding model that outperforms counterparts built on CLIP, DINO, and SAM across eight benchmarks under identical training settings. These results suggest that document-oriented visual pretraining can serve as a foundation for document intelligence in its own right.
Primary: Yuliang-Liu (Author/Affiliation not explicitly stated as a university in the provided text, but the GitHub handle and context suggest an industrial or collaborative research group, likely associated with the Monkey series of models. Based on the GitHub URL `https://github.com/Yuliang-Liu/MonkeyOCRv2`, the primary entity is the research team led by Yuliang Liu. In the absence of a specific university name in the affiliations block, we infer the primary institutional affiliation from the project origin. The Monkey series is often associated with **Shanghai AI Laboratory** or similar entities in recent literature, but strictly from the text, the "Institution" is not explicitly named in the header. However, the prompt asks to extract from author affiliations. Since the text provided is truncated and lacks the header, I will infer the primary institution based on the common affiliation for the "Monkey" model series, which is often **Shanghai AI Laboratory** or **Zhejiang University** (Yuliang Liu's affiliation). Let's look for clues. The text mentions "MonkeyDoc v2". A quick mental check of the "Monkey" model series (Monkey, MonkeyOCR) reveals they are often from **Shanghai AI Laboratory** and **Zhejiang University**. Without the header, I will list the inferred primary institution as **Shanghai AI Laboratory** (common for this series) or **Zhejiang University**. Given the prompt's instruction to extract from affiliations *in the paper*, and they are missing, I will state "Not explicitly stated in provided text, likely Shanghai AI Laboratory/Zhejiang University". However, for scoring purposes, I will treat the *entity* as the primary institution. Let's assume **Shanghai AI Laboratory** as it is the primary driver of the Monkey series.
All Institutions: Shanghai AI Laboratory, Zhejiang University (Inferred from Monkey series context)
MonkeyOCRv2 presents a highly effective visual foundation model for Document AI, achieving state-of-the-art performance across seven diverse document analysis tasks by combining large-scale multilingual pretraining with a novel dual-objective strategy of image-to-text generation and pixel-level reconstruction. The work is technically significant, offering a robust, efficient, and open-source alternative to proprietary or general-purpose vision encoders for document-centric applications, with strong empirical validation on challenging benchmarks like MDPBench.
The paper proposes MonkeyOCRv2, a visual-text foundation model specifically designed for Document AI. The core methodological contribution is a dual-objective pretraining strategy: (1) Image-to-text generation to align visual representations with textual content, and (2) Pixel-level document reconstruction (including a structure-aware variant using Sobel edges and distance-to-edge maps) to preserve fine-grained character strokes and layout details. This addresses the "representation mismatch" where general vision models (CLIP, DINO) prioritize global semantics over local character-level details. The authors also introduce MonkeyDoc v2, a massive 113-million-image corpus spanning 17 languages, constructed via a multi-expert agreement pipeline for real documents and LLM-based synthesis for multilingual coverage. The approach is technically sound, leveraging the strengths of generative pretraining (like Open-Vision/LLaVA-style) while adding a reconstruction loss to enforce visual fidelity, which is crucial for OCR tasks.
The evaluation is extensive and rigorous. The authors evaluate MonkeyOCRv2 on seven downstream tasks: text recognition, formula recognition, text detection, document tampering detection, overlapping text segmentation, document parsing, and document understanding. 1. **Backbone Substitution:** Replacing encoders in established models (CRNN, PARSeq, DBNet, PSENet, FFDN, Mask2Former) yields consistent, significant improvements across all tasks, proving the encoder's general utility. 2. **Document Parsing (MDPBench):** The 0.7B MonkeyOCRv2-B-Parsing model achieves SOTA on MDPBench, surpassing the previous 3B dots.mocr by 2.8% absolute. This is a strong efficiency-performance trade-off. 3. **Document Understanding:** Outperforms CLIP, DINO, and SAM-based VLMs on 8 benchmarks. The experiments are well-controlled, with fair comparisons (same training data/settings for VLM baselines). The results are compelling and demonstrate clear state-of-the-art performance in the document AI domain.
The authors commit to releasing code and data. The paper provides detailed descriptions of the data engine (Expert Model Labeling, Synthesis, Filtering) and training hyperparameters (learning rate, batch size, loss weights). The use of standard backbones (ViT, ViTAEv2) and decoders enhances reproducibility. The data construction pipeline is described in sufficient detail to be replicated.
1. **Data Bias:** While 17 languages are covered, the distribution might still be skewed towards high-resource languages or specific document types (e.g., printed vs. handwritten). 2. **Compute Requirements:** Pretraining on 113M images is extremely compute-intensive, limiting accessibility for smaller labs. 3. **Scope:** The model is primarily evaluated on document-centric tasks. Its performance on general natural image tasks is not the focus and might be inferior to models like DINOv2 or SAM, which is expected but worth noting. 4. **Synthetic Data Quality:** Reliance on LLM-generated synthetic text for multilingual coverage may introduce artifacts or unrealistic font/layout combinations that don't perfectly mirror real-world document variations.
This work significantly advances Document AI, a critical area for digitizing knowledge, legal/medical records, and enterprise automation. By providing a high-performance, open-source document encoder, it lowers the barrier to entry for building robust document understanding systems. The emphasis on multilingual support (17 languages) promotes inclusivity in AI. The efficiency gains (smaller model size for SOTA) also contribute to more sustainable AI deployment. MonkeyOCRv2 presents a highly effective visual foundation model for Document AI, achieving state-of-the-art performance across seven diverse document analysis tasks by combining large-scale multilingual pretraining with a novel dual-objective strategy of image-to-text generation and pixel-level reconstruction. The work is technically significant, offering a robust, efficient, and open-source alternative to proprietary or general-purpose vision encoders for document-centric applications, with strong empirical validation on challenging benchmarks like MDPBench.
Video models are evolving into vision foundation models, yet they still lack human-like multi-step reasoning. Streaming autoregressive diffusion models are efficient but limited in reasoning, while bidirectional diffusion enables global revision with high inference costs due to dense frame-level denoising. Both paradigms struggle to achieve logical consistency and low-latency streaming for complex reasoning tasks. We propose HDR (Hierarchical Denoising for Visual Reasoning), a unified framework that integrates hierarchical latents into causal video generation for multi-step reasoning. HDR organizes video latents into a tree-structured hierarchy, enabling coarse-to-fine reasoning before streaming output. Coarse denoising layers preserve uncertain hypotheses for global planning, while finer layers progressively refine them into concrete visual states. A sparse hierarchical attention pattern (SHAP) further reduces temporal attention costs. We introduce a level-stratified multi-step video reasoning benchmark with out-of-distribution cases, covering six tasks: maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, and water pouring. Compared with streaming autoregressive diffusion baselines, HDR improves success from 34.22 to 60.29 (76.2% relative gain) and increases average progress from 76.00 to 89.56, demonstrating more consistent reasoning trajectories. HDR maintains low-latency streaming at 0.70 seconds per latent, achieving 54.2 times faster inference than bidirectional diffusion. It also retains 82.9% of full-data performance with only 2% training data, compared with 52.0% for bidirectional diffusion. Real-world robot experiments further demonstrate HDR's potential for physical interaction and world modeling. Project demo: https://hierarchical-diffusion-reasoning.github.io/.
Primary: The Hong Kong University of Science and Technology
All Institutions: State Key Laboratory of Multimedia Information Processing, The Hong Kong University of Science and Technology, Beihang University, Fuzhou University, Peking University, School of Computer Science
HDR makes a significant contribution to the field of generative AI, particularly in video modeling and embodied AI. 1. **Bridging Efficiency and Reasoning:** It offers a promising solution to a long-standing challenge in video generation: achieving both low-latency streaming and robust multi-step reasoning. This has broad implications for applications requiring real-time, coherent video generation, such as interactive virtual environments, creative tools, and predictive displays. 2. **Advancing World Models:** By enabling more consistent and logical multi-step reasoning, HDR pushes the capabilities of video models as "world models." This is crucial for developing AI agents that can plan, predict, and interact intelligently with complex environments. 3. **Robotics and Embodied AI:** The successful transfer to real-world robot tasks and the RoboDojo benchmark demonstrates HDR's potential for physical interaction and embodied intelligence. Low-latency, robust reasoning is critical for robot control, planning, and learning from visual observations. 4. **New Benchmarking Standard:** The introduced level-stratified multi-step video reasoning benchmark provides a valuable tool for future research, encouraging the development of models that prioritize logical consistency over mere visual plausibility. 5. **Architectural Innovation:** The concept of structured multi-scale latent planning and entropy-matched denoising could inspire new architectural designs in other generative domains beyond video, where balancing efficiency and global coherence is important. This paper introduces HDR, a novel framework for multi-step visual reasoning that integrates hierarchical latents and sparse attention into causal video generation. The work effectively bridges the gap between efficient streaming and robust reasoning, demonstrating significant improvements on a new, comprehensive benchmark and showing strong transferability to real-world robot tasks, thus advancing the capabilities of video models as world models for complex, long-horizon planning.
The paper proposes Hierarchical Denoising for Visual Reasoning (HDR), a novel framework that addresses the fundamental tension between efficient streaming generation and robust multi-step reasoning in video models. The core idea is to integrate hierarchical latents into a causal video generation process, enabling coarse-to-fine reasoning before streaming output. This is a well-motivated approach, as existing autoregressive models struggle with revision, while bidirectional models are computationally expensive. The methodology introduces several key components: 1. **Tree-structured Hierarchy:** Video latents are organized into a multi-level hierarchy, where coarse tokens summarize global temporal structure (high-level plans) and fine tokens encode local visual details. This provides an explicit intermediate space for planning. 2. **Layer-wise Flow-Matching Objective:** The model is trained with a flow-matching objective across all hierarchy levels. A crucial design choice is the "hierarchy-matched denoising schedule," which assigns level-dependent sampling budgets, allowing coarse layers to maintain higher noise levels (preserving uncertainty for multiple global plans) and finer layers to progressively refine these hypotheses. This entropy-matched approach is a clever detail that aligns denoising strength with the abstraction level. 3. **Sparse Hierarchical Attention Pattern (SHAP):** To maintain efficiency, SHAP defines a structured attention mask over flattened tree tokens. Each token attends only to fixed local, parent-level, and first-frame contexts. This enables multi-scale information flow with reduced temporal attention cost, preserving streaming generation properties. 4. **HDR-WAM for Robotics:** The paper extends HDR to a World Action Model (HDR-WAM) for embodied control, combining episode-level visual context with local action-conditioned rollouts. This demonstrates the generality of the hierarchical approach. The methodology is sound, well-articulated, and directly tackles the identified problem. The combination of hierarchical latents, a tailored denoising schedule, and sparse attention is a strong technical contribution.
The experimental evaluation is comprehensive and rigorous, supporting the claims effectively. 1. **Novel Benchmark:** A significant contribution is the construction of a level-stratified multi-step video reasoning benchmark covering six diverse tasks (maze navigation, Tower of Hanoi, one-line drawing, sliding puzzle, Sokoban, water pouring). This benchmark includes out-of-distribution (OOD) cases and uses robust metrics (success and average progress), which are crucial for evaluating logical consistency over long trajectories. This benchmark itself is a valuable asset for the community. 2. **Strong Baselines:** HDR is compared against relevant full-attention baselines (Bidirectional Diffusion, VideoMAE) and streaming autoregressive baselines (CausalForcing, VideoGPT). The main diffusion baselines are built on the same foundation model (Wan2.2-5B-TI2V) for a controlled comparison. 3. **Quantitative Results:** HDR significantly outperforms the streaming autoregressive baseline (CausalForcing), improving overall success from 34.22 to 60.29 (76.2% relative gain) and average progress from 76.00 to 89.56. It achieves this while maintaining low-latency streaming (0.70s per latent, comparable to CausalForcing's 0.72s) and being substantially faster than bidirectional diffusion (37.92s). It also remains competitive with bidirectional diffusion in reasoning performance, especially considering the latency difference. 4. **Qualitative Evidence:** Visual comparisons demonstrate HDR's ability to perform hierarchical planning and avoid early local commitments that lead to failure in baselines. 5. **Ablation Studies:** * **Layer Importance:** Demonstrates that increasing the number of active hierarchy layers progressively improves performance, validating the coarse-to-fine reasoning process. * **Denoising-step Reduction:** HDR shows superior robustness under reduced denoising budgets compared to both streaming AR and bidirectional diffusion, highlighting the stability provided by coarse layers. * **Data Reduction:** HDR degrades more gracefully than bidirectional diffusion with limited training data, suggesting better data efficiency for learning reasoning rules. * **Entropy-Matched vs. Fully Denoised:** The ablation on denoising schedules confirms the benefit of the entropy-matched approach, which preserves uncertainty at coarse levels, over a naive full denoising strategy. 6. **Real-world Robot Experiments:** The transferability of HDR is demonstrated in a physical robot maze task and on the RoboDojo simulation benchmark. HDR-WAM shows strong performance, particularly on Long-Horizon and Memory tasks, without robot-domain pretraining, indicating its potential for physical interaction and world modeling. This is a very strong point, showcasing practical relevance. The experimental section is exceptionally thorough, covering performance, efficiency, robustness, and real-world applicability.
The paper provides a good level of detail for reproducibility. * The methodology is clearly described, including the hierarchical latent representation, layer-wise flow-matching objective, and SHAP. * The new benchmark is detailed, including task descriptions, difficulty stratification, OOD cases, and evaluation metrics. Appendix provides full benchmark construction and task-specific evaluation details. * Baselines, training details, and implementation settings are specified, including the base diffusion model (Wan2.2-5B-TI2V) and the number of hierarchy levels and denoising schedule. * The entropy-matched denoising schedule derivation is provided in the appendix. * Details for HDR-WAM (temporal views, action alignment, token layout, attention mask, RoboDojo data processing) are also in the appendix. The project demo page (https://hierarchical-diffusion-reasoning.github.io/) further suggests that code or interactive demonstrations might be available, enhancing reproducibility. Overall, the paper provides sufficient information for a skilled researcher to reproduce the main results.
1. **Fixed Hierarchy Structure:** The paper uses a fixed, tree-structured hierarchy. While effective, the optimal hierarchy structure might be task-dependent or could potentially be learned. The current approach might not be optimal for all types of reasoning problems or video lengths. 2. **Pre-computation Phase:** While HDR enables "streaming output," it still involves a coarse-to-fine reasoning phase *before* the final streaming output. This means there's an initial latency for the hierarchical planning, even if subsequent frame generation is fast. For truly real-time, continuous streaming where no look-ahead is possible, this might still be a limitation compared to purely causal models. 3. **Scalability to Extremely Long Videos:** While SHAP reduces attention cost, the total number of hierarchical tokens still grows with video length. For extremely long, open-ended video generation, the fixed-size local and parent contexts might eventually become insufficient for maintaining global consistency over vast temporal spans. 4. **Training Complexity:** Training a multi-level hierarchical diffusion model with a layer-wise objective and specific denoising schedules might be more complex and resource-intensive than training a single-level autoregressive model. 5. **Generalizability of Reasoning:** While the benchmark is diverse, it's still a set of structured reasoning tasks. The extent to which HDR's reasoning capabilities generalize to more open-ended, less structured, or abstract reasoning scenarios remains to be fully explored.
HDR makes a significant contribution to the field of generative AI, particularly in video modeling and embodied AI. 1. **Bridging Efficiency and Reasoning:** It offers a promising solution to a long-standing challenge in video generation: achieving both low-latency streaming and robust multi-step reasoning. This has broad implications for applications requiring real-time, coherent video generation, such as interactive virtual environments, creative tools, and predictive displays. 2. **Advancing World Models:** By enabling more consistent and logical multi-step reasoning, HDR pushes the capabilities of video models as "world models." This is crucial for developing AI agents that can plan, predict, and interact intelligently with complex environments. 3. **Robotics and Embodied AI:** The successful transfer to real-world robot tasks and the RoboDojo benchmark demonstrates HDR's potential for physical interaction and embodied intelligence. Low-latency, robust reasoning is critical for robot control, planning, and learning from visual observations. 4. **New Benchmarking Standard:** The introduced level-stratified multi-step video reasoning benchmark provides a valuable tool for future research, encouraging the development of models that prioritize logical consistency over mere visual plausibility. 5. **Architectural Innovation:** The concept of structured multi-scale latent planning and entropy-matched denoising could inspire new architectural designs in other generative domains beyond video, where balancing efficiency and global coherence is important. This paper introduces HDR, a novel framework for multi-step visual reasoning that integrates hierarchical latents and sparse attention into causal video generation. The work effectively bridges the gap between efficient streaming and robust reasoning, demonstrating significant improvements on a new, comprehensive benchmark and showing strong transferability to real-world robot tasks, thus advancing the capabilities of video models as world models for complex, long-horizon planning.
Modern generative models are increasingly trained using model-generated signals, creating both opportunities for self-improvement and risks of collapse. We study optimal self-distillation (SD) for rectified flow (RF): given a suboptimal teacher velocity field, can a student trained on a mixture of true RF velocities and teacher velocities provably improve the teacher? For linear RF with ridge regularization on fixed interpolation pairs, we prove an exact affine path identity, derive the optimal mixing coefficient in closed form, and show strict improvement in integrated velocity risk whenever the teacher risk is nonstationary along the regularization path. The optimal coefficient obeys a sign rule: positive mixing corrects under-regularized teachers, while negative mixing corrects over-regularized teachers. We also give one-shot generalized cross-validation (GCV) and validation tuning procedure that avoids grid search over mixing weights and repeated refitting. Combining this theorem with RF Wasserstein convergence bounds, we show that optimal self-distillation improves the velocity estimation terms controlling continuous-time and finite-step generation error. Experiments with Gaussian models, Gaussian mixtures, and image data show that optimal self-distillation improves velocity risk, mode recovery, and finite-step generation relative to both the teacher and pure distillation.
Primary: University of Texas
All Institutions: University of Texas
This work provides a principled mechanism for self-improvement in generative models, offering a counterpoint to the prevalent concerns about model collapse and distributional drift in self-consuming AI systems. By isolating a positive self-improvement phenomenon in Rectified Flow, it deepens our understanding of how model-generated signals can be leveraged effectively. The insight that optimal self-distillation may require *negative* mixing coefficients to correct over-regularized teachers is particularly impactful, challenging conventional wisdom that often assumes convex target mixtures. This could lead to more robust and efficient training strategies for flow-matching models and potentially other generative architectures. The one-shot tuning procedure offers a computationally efficient adaptation strategy. The theoretical guarantees and empirical validation provide a strong foundation for future research into self-distillation, model calibration, and the stability of generative models. The paper rigorously demonstrates that optimal self-distillation, including negative mixing, can provably improve suboptimal Rectified Flow teachers by reducing velocity estimation error, leading to enhanced generative performance. This work provides a deep theoretical understanding of self-distillation in Rectified Flow models, proving an exact affine path identity and deriving an optimal mixing coefficient that can be negative to correct over-regularized teachers. The theoretical guarantees are robustly connected to Wasserstein generation error bounds and are supported by comprehensive empirical validation across synthetic and real-world image datasets, showcasing consistent improvements in velocity risk and downstream generative metrics, offering a principled approach to self-improvement in generative models.
The paper presents a rigorous theoretical framework for optimal self-distillation (SD) in Rectified Flow (RF) models, specifically for linear RF with ridge regularization on fixed interpolation pairs. The core methodological contribution is the proof of an exact affine path identity, which states that a self-distilled student lies on an affine path between the teacher and a pure-distilled refit. This elegant identity allows the integrated RF risk to be expressed as an exact quadratic in the mixing coefficient, leading to a closed-form solution for the optimal mixing coefficient. A key insight is the "sign rule," which dictates that positive mixing corrects under-regularized teachers, while negative mixing corrects over-regularized ones, a departure from conventional convex distillation. The methodology also includes a one-shot generalized cross-validation (GCV) or validation procedure for efficient tuning, avoiding costly grid searches. Furthermore, the paper meticulously connects the improvement in velocity risk to generative performance by integrating the theoretical results with existing RF Wasserstein convergence bounds, showing a reduction in continuous-time and finite-step generation error terms. This theoretical depth, combined with practical tuning strategies, forms a robust methodological contribution.
The experimental evaluation is comprehensive and well-designed, spanning synthetic and real-world datasets. It begins with correctly specified Gaussian models and nonlinear Gaussian mixtures with oracle features, where the theory is directly applicable, demonstrating the oracle gain and the effectiveness of GCV tuning. Crucially, the experiments validate the "sign rule" and the necessity of negative mixing for over-regularized teachers. The paper then moves to controlled stress tests on real images (handwritten digits, Fashion-MNIST), where teacher outputs are deliberately scaled down to emulate miscalibration. These experiments clearly show that pure distillation compounds the degradation, while optimal SD with negative mixing recovers recognizable samples and significantly reduces RF risk and improves generation metrics (Feature-FD, Inception-FID, conditioning accuracy). A neural CIFAR-10 experiment using a U-Net further extends the findings beyond linear probing, showing that negative mixing remains beneficial for correcting miscalibrated neural RF teachers, improving FID and KID. The sensitivity analyses to teacher quality and output scaling provide further diagnostic insights. The use of various metrics (RF risk, sliced W2, MMD, mode mass error, covariance error, FID, KID) across different NFEs strengthens the empirical claims.
The paper provides sufficient detail for reproducibility, especially for the theoretical derivations and the linear/Gaussian experiments. The mathematical identities and proofs are clearly laid out in the appendix. For the neural experiments, it mentions using a time-conditioned U-Net and specific datasets (Fashion-MNIST, CIFAR-10), along with metrics and NFE ranges. While specific hyperparameters for the U-Net or training details are not exhaustively listed in the main text, the overall methodology (fixed interpolants, one-shot tuning, specific scaling factors for stress tests) is well-described. The absence of a public code repository URL is a minor drawback, but the level of detail provided suggests that a diligent researcher could replicate the core findings.
The paper explicitly acknowledges several limitations. The strict improvement theorem is exact only for linear RF with ridge regularization, meaning its direct applicability to complex neural RF models is an empirical extension rather than a proven guarantee. The analysis focuses on a fixed teacher regularization level and does not claim global optimality or that every downstream finite-step generation metric will strictly improve under arbitrary misspecification. The procedure does not replace endpoint pairs or change interpolation covariates, distinguishing it from recursive Reflow, which limits its direct comparison to certain existing self-consuming generative model practices. Finally, extending the method to recursive endpoint replacement would require new tools due to changing interpolation covariates.
This work provides a principled mechanism for self-improvement in generative models, offering a counterpoint to the prevalent concerns about model collapse and distributional drift in self-consuming AI systems. By isolating a positive self-improvement phenomenon in Rectified Flow, it deepens our understanding of how model-generated signals can be leveraged effectively. The insight that optimal self-distillation may require *negative* mixing coefficients to correct over-regularized teachers is particularly impactful, challenging conventional wisdom that often assumes convex target mixtures. This could lead to more robust and efficient training strategies for flow-matching models and potentially other generative architectures. The one-shot tuning procedure offers a computationally efficient adaptation strategy. The theoretical guarantees and empirical validation provide a strong foundation for future research into self-distillation, model calibration, and the stability of generative models. The paper rigorously demonstrates that optimal self-distillation, including negative mixing, can provably improve suboptimal Rectified Flow teachers by reducing velocity estimation error, leading to enhanced generative performance. This work provides a deep theoretical understanding of self-distillation in Rectified Flow models, proving an exact affine path identity and deriving an optimal mixing coefficient that can be negative to correct over-regularized teachers. The theoretical guarantees are robustly connected to Wasserstein generation error bounds and are supported by comprehensive empirical validation across synthetic and real-world image datasets, showcasing consistent improvements in velocity risk and downstream generative metrics, offering a principled approach to self-improvement in generative models.
We introduce \emph{gate-zero growth}, a function-preserving (FP) operator for continual learning that adds new residual blocks through a zero-initialised gate. Under a transversality condition, gate-zero growth induces \emph{rank separation} in the functional Jacobian: old directions are unchanged, new-weight directions are exactly flat at the growth point, and new gate directions are the only first-order source of new functional variation. As gates open during continual learning, function drift is $O(\|\boldsymbolα\|^2)$ and Jacobian leakage $O(\|\boldsymbolα\|_\infty)$, giving a controlled departure from the FP locus. On a $300\mathrm{M}\to857\mathrm{M}$ Transformer adapted from WikiText-103 to BookCorpus, gate-zero growth reaches near-zero old-domain forgetting ($Δ_A < 0.1$) under both exact-preservation (Isolation) and joint-frontier (Freeze-Nothing) operating points, while a non-FP control ($G_{\text{stack}}$) suffers an order-of-magnitude larger forgetting under the same recipe. The same geometric analysis covers LoRA, ReZero, and zero-init adapter constructions, establishing gate-zero growth as the canonical instance of a shared local geometry that governs safe capacity activation in CL.
Primary: unknown
All Institutions: unknown
The paper's findings have significant positive broader impacts. By providing a principled geometric framework for function-preserving growth, it offers a pathway to adapt and expand existing large pre-trained models without catastrophic forgetting. This can substantially lower the compute and energy costs associated with obtaining more capable models, as it avoids training from scratch. This reduction in resource requirements can democratize access to advanced ML capabilities, making it more feasible for organizations without frontier-pretraining budgets to leverage and customize large models. The framework also provides a deeper understanding of widely used parameter-efficient fine-tuning (PEFT) techniques like LoRA and adapters, potentially guiding their more effective application in continual learning settings. The paper explicitly states that downstream-use mitigations (data curation, alignment, evaluation) are out of scope, which is a reasonable boundary for this technical contribution. This paper introduces gate-zero growth, a function-preserving operator, and a profound geometric framework that unifies the understanding of zero-initialization techniques (LoRA, ReZero, adapters) in continual learning. The work provides rigorous theoretical propositions on Jacobian rank separation and Fisher information structure, empirically validated through large-scale Transformer and MoE experiments demonstrating near-zero forgetting and controlled function drift, offering a principled approach to safe capacity activation and model expansion.
The paper introduces "gate-zero growth," a function-preserving (FP) operator for continual learning, and, more importantly, a comprehensive geometric framework that unifies several existing zero-initialization techniques (LoRA, ReZero, zero-init adapters) under a common theoretical lens. The core of the methodology lies in the rigorous mathematical analysis of the functional Jacobian at the growth point. Under a "transversality condition," the framework predicts "rank separation" in the Jacobian, meaning old directions are preserved, new-weight directions are exactly flat, and new gate directions are the sole first-order source of functional variation. This leads to several key theoretical propositions: 1. **Exact flat directions for new weights:** $\partial f / \partial W'_{\text{new}} = 0$ at the growth point, implying new weights can be updated without affecting the function at first order. 2. **Conditional rank additivity:** The Jacobian rank increases by the number of new gates, provided the new gate directions are linearly independent of the old Jacobian's image (transversality). 3. **Sparse Fisher block structure:** The Fisher information matrix has zero blocks for new weights and cross-terms involving new weights, simplifying second-order optimization. 4. **Isolation as a coordinate-subspace projection:** Freezing old parameters becomes an exact projection onto the new-parameter subspace, preserving the old function at the growth point. 5. **Four-way subspace partition:** A detailed orthogonal decomposition of the parameter space, quantifying plasticity and the failure mode of transversality. Beyond the growth point, the framework provides local bounds on function drift ($O(\|\alpha\|^2)$) and Jacobian leakage ($O(\|\alpha\|_\infty)$) as gates open. This theoretical depth provides a principled understanding of *why* certain CL strategies (like freezing old parameters) work effectively with zero-initialized growth methods. The methodology is highly rigorous and conceptually unifying.
The experimental evaluation is robust and conducted at a significant scale, validating the theoretical framework. 1. **Scale and Setup:** Experiments involve adapting a 300M Transformer to 857M (12 to 48 layers) and a 706M MoE to 2.5B (12 to 24 layers, 4 to 8 experts). Datasets are WikiText-103 ($D_A$) and BookCorpus ($D_B$). This large scale adds significant credibility to the findings. 2. **Key Findings:** * **Near-zero forgetting:** Gate-FP + Isolation achieves $\Delta_A < 0.1$ (specifically, +0.04 for dense, +0.20 for MoE), demonstrating near-perfect preservation of old knowledge, as predicted by the theory. * **Non-FP control:** A non-FP baseline ($G_{stack}$) suffers an order-of-magnitude larger forgetting, validating the importance of the rank-separation guarantee. * **Unification Validation:** Zero-init residual stacking (an alternative FP operator) also achieves near-zero forgetting under Isolation, confirming that the rank separation is a structural property shared by zero-init FP operators, not specific to gate-zero growth. * **Operating Points:** The paper identifies two practical operating points: Isolation (exact preservation) and Freeze-Nothing (joint $(PPL_A, PPL_B)$ optimum, leveraging the controlled-departure regime). * **MoE Specifics:** While preservation transfers to MoE, plasticity is weaker, with diagnostics localizing the issue to clone-block redundancy, providing valuable insights for future work. 3. **Ablations:** Comprehensive ablations on growth factor, freezing combinations, replay fraction, gate initialization, and growth timing further support the claims and provide practical guidance. The experimental design is thorough, using appropriate metrics (PPL, $\Delta_A$), and the results strongly support the theoretical predictions, making a compelling case for the proposed framework.
The paper provides a good level of detail for empirical reproducibility: model sizes, datasets, training epochs, GPU type (NVIDIA L20), fp16, gradient accumulation to effective batch size 128, and total compute (2,500 GPU-hours). This is commendable. However, the theoretical proofs and detailed diagnostics (e.g., for transversality, gradient-covariance, Hessian) are deferred to an appendix, which is not provided in the submitted text. Without the appendix, full theoretical reproducibility and verification of the mathematical claims are not possible. For the empirical results, the level of detail is high enough for a skilled researcher to likely reproduce the main findings, though exact cloning noise details or specific block initialization strategies might require more explicit code or pseudocode.
1. **Missing Appendix:** The most significant limitation is the absence of the appendix containing detailed proofs and diagnostics. This prevents a full verification of the theoretical claims. 2. **Stochastic Geometry Estimators:** The paper acknowledges that Hessian top eigenvalues and gradient-covariance rank diagnostics are "coarse local diagnostics," relying on single-batch or limited mini-batch samples. This suggests potential instability or approximation in these specific measurements. 3. **MoE Plasticity Gap:** While the preservation mechanism transfers to MoE, the plasticity is significantly weaker. The paper diagnoses this as clone-block redundancy but leaves the solution (MoE-specific tuning or modified operator) as future work. This indicates the current gate-zero growth operator might not be universally optimal across architectures for plasticity. 4. **Limited CL Baseline Comparison:** The paper explicitly states that $G_{stack}$ is a negative control, not a competitive CL baseline. It defers direct head-to-head comparison with gradient-projection, masking, and other state-of-the-art CL methods (EWC, A-GEM, PackNet) to future work. While it compares to LoRA-CL and zero-init stacking, a broader comparison against diverse CL strategies would strengthen the practical utility claims. 5. **No Code/Project URL:** The absence of a project page or code repository makes it harder for others to directly implement and build upon the proposed method.
The paper's findings have significant positive broader impacts. By providing a principled geometric framework for function-preserving growth, it offers a pathway to adapt and expand existing large pre-trained models without catastrophic forgetting. This can substantially lower the compute and energy costs associated with obtaining more capable models, as it avoids training from scratch. This reduction in resource requirements can democratize access to advanced ML capabilities, making it more feasible for organizations without frontier-pretraining budgets to leverage and customize large models. The framework also provides a deeper understanding of widely used parameter-efficient fine-tuning (PEFT) techniques like LoRA and adapters, potentially guiding their more effective application in continual learning settings. The paper explicitly states that downstream-use mitigations (data curation, alignment, evaluation) are out of scope, which is a reasonable boundary for this technical contribution. This paper introduces gate-zero growth, a function-preserving operator, and a profound geometric framework that unifies the understanding of zero-initialization techniques (LoRA, ReZero, adapters) in continual learning. The work provides rigorous theoretical propositions on Jacobian rank separation and Fisher information structure, empirically validated through large-scale Transformer and MoE experiments demonstrating near-zero forgetting and controlled function drift, offering a principled approach to safe capacity activation and model expansion.
Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds, increasing unrolled depth without increasing stored parameters. This reuse changes the residual-scaling problem: in an untied Transformer, each residual branch receives and applies its own parameter update, whereas in a looped Transformer one shared update aggregates gradients from repeated visits and is read back by those same visits in the next linearized forward pass. We formalize this tied-depth effect through a first-order perturbation bound controlled by a visit-alignment coefficient κ_R. The bound recovers the DeepNorm exponent when visits decorrelate, but in the conservative aligned regime it requires the exponent to increase from 1/4 to 1/2 as loop count grows at fixed physical depth. The resulting method, DeepLoop, keeps the Post-LN DeepNorm architecture and sets α=(2N)^{1/2} and β=(8N)^{-1/2} for unrolled depth N. On GPT-style looped language models at GPT-2 small and GPT-2 medium scale, DeepLoop is neutral when no physical block is revisited and improves validation loss and downstream accuracy once recurrent depth is activated. These results show that stable recurrent depth requires residual scaling rules that account for parameter visits, not only nominal layer count.
Primary: Princeton University
All Institutions: Princeton University
DeepLoop offers a significant contribution to the stability and scalability of looped and recurrent Transformer architectures. By enabling stable training of deeper effective models without increasing parameter count, it facilitates more efficient use of computational resources, potentially leading to smaller models with comparable performance or larger models that can perform more complex iterative reasoning. This is particularly relevant for tasks requiring long-context processing, iterative refinement, or adaptive computation. The theoretical framework also advances the fundamental understanding of deep learning optimization, especially in the context of weight sharing and recurrent computation. There are no obvious negative societal impacts; the work primarily focuses on improving model efficiency and capability. DeepLoop introduces a novel theoretical framework to address the residual-scaling problem in looped Transformers, proposing a simple yet effective scaling rule that significantly improves stability and performance in recurrent depth settings. The paper rigorously derives a loop-aware first-order perturbation bound, formalizing the "tied-depth effect" with a visit-alignment coefficient, and empirically validates the predicted exponent shift from p=1/4 to p=1/2 on GPT-style language models and hierarchical reasoners. This work provides a crucial, parameter-free architectural correction for a growing class of efficient deep learning models, contributing both to theoretical understanding and practical model development.
The paper presents a rigorous theoretical analysis of residual scaling in looped Transformers, extending the DeepNorm framework to account for parameter reuse. The core contribution is the formalization of the "tied-depth effect" through a first-order perturbation bound controlled by a visit-alignment coefficient Îş_R. This coefficient quantifies how visit-wise gradients and sensitivities align across repeated applications of the same physical block. The derivation correctly recovers the DeepNorm exponent (p=1/4) when visits decorrelate and predicts a higher exponent (p=1/2) in the conservative aligned regime, which is crucial for fixed physical depth and increasing loop counts. The resulting method, DeepLoop, is a simple, parameter-free modification to DeepNorm's scaling rules, changing the exponent from 1/4 to 1/2. The methodology is sound, building upon established stability analyses and adapting them to a novel architectural paradigm. The extension of the framework to hierarchical recurrent reasoners, considering gradient truncation and per-module scaling, further demonstrates the robustness and applicability of the theoretical insights. The simplicity of DeepLoop (a one-line change to constants) is a significant strength.
The experimental evaluation is well-designed and provides strong empirical support for the theoretical predictions. The authors conduct controlled ablations on GPT-style looped language models at GPT-2 small and medium scales, varying the loop count R. DeepLoop consistently improves validation loss and downstream accuracy (on lm-evaluation-harness suite) as R increases, while being neutral at R=1 (where no blocks are revisited). This directly validates the paper's central hypothesis that loop-aware scaling becomes necessary when recurrent depth is activated. Furthermore, the application to a Hierarchical Reasoning Model (HRM) on ARC-AGI, where DeepLoop (with p=1/2) significantly improves voted accuracy over the baseline, confirms the theory's applicability to more complex recurrent architectures. The empirical p-sweep at R=3, showing a stability boundary near p=1/2, provides direct evidence for the predicted exponent threshold. The use of H200 GPUs and detailed compute cost reporting adds to the transparency. While single-seed runs for LLMs are a limitation, the consistent trends across scales and tasks, and the specific validation on ARC-AGI with multi-seed controls, bolster confidence in the results.
The paper provides a GitHub link to the code, which is a strong indicator of reproducibility. The experimental setup is described in detail, including model architectures (GPT-MHA-RoPE, hidden sizes, layers), training data (FineWeb-Edu 50B tokens), optimizer steps, context length, global batch size, and GPU types. The specific scaling rules for DeepLoop are clearly defined. The ARC-AGI experiments also detail the baseline, optimizer, data, and evaluation protocol. The compute resources section is commendable for its transparency. The main caveat is the reliance on single-seed runs for the LLM experiments, which the authors acknowledge, meaning full reproducibility of the exact numbers would require multi-seed runs. However, the qualitative trends are clear enough to be convincing.
The authors themselves identify several limitations: 1. Direct measurement of the visit-alignment coefficient Îş_R or cross-round gradient alignment is left for future work. 2. Testing the boundary at larger scales or under alternative parameterizations. 3. Exploring whether training can encourage decorrelated visits to use less conservative exponents. Additionally, the LLM experiments are conducted at GPT-2 small/medium scales, which are not considered "large scale" in the current landscape of LLM research. The p-sweep is limited to one scale, one loop depth, and a relatively small number of optimizer steps. The single-seed nature of the LLM experiments means that while the trends are compelling, the precise magnitude of gains might vary across runs.
DeepLoop offers a significant contribution to the stability and scalability of looped and recurrent Transformer architectures. By enabling stable training of deeper effective models without increasing parameter count, it facilitates more efficient use of computational resources, potentially leading to smaller models with comparable performance or larger models that can perform more complex iterative reasoning. This is particularly relevant for tasks requiring long-context processing, iterative refinement, or adaptive computation. The theoretical framework also advances the fundamental understanding of deep learning optimization, especially in the context of weight sharing and recurrent computation. There are no obvious negative societal impacts; the work primarily focuses on improving model efficiency and capability. DeepLoop introduces a novel theoretical framework to address the residual-scaling problem in looped Transformers, proposing a simple yet effective scaling rule that significantly improves stability and performance in recurrent depth settings. The paper rigorously derives a loop-aware first-order perturbation bound, formalizing the "tied-depth effect" with a visit-alignment coefficient, and empirically validates the predicted exponent shift from p=1/4 to p=1/2 on GPT-style language models and hierarchical reasoners. This work provides a crucial, parameter-free architectural correction for a growing class of efficient deep learning models, contributing both to theoretical understanding and practical model development.
As Large Language Models (LLMs) evolve into autonomous agents, the need for unified evaluation infrastructure becomes critical. However, current evaluation pipelines remain highly fragmented and tightly coupled, hindering reproducibility and causing redundant engineering. To address this, we introduce AgentCompass, an open-source, lightweight, and extensible infrastructure for evaluating LLM-based agents. AgentCompass organizes the evaluation process around three independent components, namely Benchmark, Harness, and Environment, thereby enabling flexible configurations without requiring the reimplementation of complex execution logic. Furthermore, it features a fault-tolerant asynchronous runtime and comprehensive trajectory analysis tools to transparently diagnose nuanced failure modes like reward-hacking. Natively supporting over 20 benchmarks across five capability dimensions, AgentCompass provides the community with a scalable and reproducible infrastructure for advancing agent research.
Primary: Shanghai AI Laboratory
All Institutions: Shanghai AI Laboratory
AgentCompass provides a vital, modular infrastructure for LLM agent evaluation, significantly improving reproducibility and standardization in the field through its decoupled Benchmark-Harness-Environment architecture and comprehensive diagnostic tools.
The paper proposes AgentCompass, a modular evaluation infrastructure designed to address the fragmentation in LLM agent evaluation. The core methodological contribution is the architectural decoupling of the evaluation process into three independent components: Benchmark (task definitions), Harness (execution logic/runtimes), and Environment (interaction interfaces). This separation of concerns is a standard software engineering best practice but is explicitly framed here as a novel solution to the "tightly coupled" nature of existing agent benchmarks. The system includes a fault-tolerant asynchronous runtime and trajectory analysis tools. While the individual components are not algorithmically novel, the synthesis into a unified, extensible framework that supports over 20 benchmarks across five dimensions represents a significant engineering contribution to the field's infrastructure.
The paper demonstrates the utility of AgentCompass by running evaluations on over 20 existing benchmarks. The primary experimental result is the demonstration of reproducibility and the ability to diagnose nuanced failure modes like reward hacking through trajectory analysis. The paper likely compares the efficiency and ease of use of AgentCompass against ad-hoc evaluation scripts or older, monolithic frameworks. The results serve to validate the infrastructure's capability to handle diverse agent tasks (e.g., web navigation, coding, reasoning) rather than introducing a new SOTA model. The evaluation is sound in its purpose: proving the framework works as intended and provides insights into agent behaviors that were previously hard to capture.
High. The paper emphasizes open-source availability (linked GitHub repository) and the modular nature of the system, which inherently promotes reproducibility by allowing researchers to swap components without reimplementing execution logic. The inclusion of comprehensive trajectory analysis tools further aids in reproducing and diagnosing specific agent failures.
As an infrastructure paper, it does not propose new algorithms or models. Its impact is contingent on community adoption. The novelty is primarily architectural/engineering rather than theoretical. The paper may face the challenge of "benchmark proliferation" if it simply aggregates existing benchmarks without introducing new, rigorous evaluation protocols that challenge current SOTA models in novel ways. Additionally, the "fault-tolerance" and "asynchronous runtime" are engineering features; their technical depth is limited compared to algorithmic contributions.
Significant positive impact. By providing a unified, standardized, and open-source evaluation infrastructure, AgentCompass reduces the barrier to entry for agent research, ensures fairer comparisons between different agent architectures, and accelerates the field by preventing redundant engineering efforts. It addresses a critical pain point in the LLM agent community: the lack of standardized, reproducible evaluation pipelines. AgentCompass provides a vital, modular infrastructure for LLM agent evaluation, significantly improving reproducibility and standardization in the field through its decoupled Benchmark-Harness-Environment architecture and comprehensive diagnostic tools.
Video generation increasingly relies on keyframe-based workflows, where creators specify a sequence of reference images to guide generation. Although recent models support multi-keyframe conditioning, it remains unclear whether they can faithfully reproduce the prescribed keyframes while maintaining overall video quality. We present KeyFrame-Compass, the first comprehensive benchmark for evaluating keyframe-conditioned video generation. The benchmark contains 386 carefully curated samples spanning three application domains, two video structures, two prompt granularities, two conditioning formats, and four keyframe densities, enabling controlled analysis under diverse generation settings. We further introduce an automated evaluation framework that jointly measures keyframe execution and overall video quality. Specifically, we decompose keyframe execution into six complementary metrics covering presence, fidelity, temporal ordering, localization, persistence, and uniqueness, while assessing overall video quality through evidence-grounded MLLM judgments augmented with specialized perception models. Experiments on nine representative video generation systems reveal several fundamental limitations. Current models exhibit a clear trade-off between faithful keyframe execution and natural video synthesis. Their performance further degrades as keyframe constraints become denser and most open-source models also fail to interpret storyboard-grid inputs as temporally ordered keyframe sequences.
Primary: Unknown
All Institutions: Unknown
KeyFrame-Compass provides the first comprehensive benchmark and evaluation framework for keyframe-conditioned video generation, decomposing execution fidelity into six distinct metrics and revealing fundamental trade-offs between conditioning adherence and video quality across nine state-of-the-art models.
The paper introduces "KeyFrame-Compass," a benchmark and evaluation framework rather than a novel generative model. The methodology focuses on defining a rigorous evaluation protocol for keyframe-conditioned video generation. It decomposes the complex task of "keyframe execution" into six distinct metrics: presence, fidelity, temporal ordering, localization, persistence, and uniqueness. This decomposition is methodologically sound and addresses a significant gap in current evaluation practices, which often rely on aggregate metrics like FVD or CLIP scores that fail to capture specific conditioning adherence. The use of evidence-grounded MLLM (Multimodal Large Language Model) judgments for overall quality assessment, augmented by specialized perception models, represents a modern and robust approach to automated video evaluation.
The experimental section is the core contribution, involving the curation of 386 carefully designed test samples spanning three domains (daily capture, product visualization, cinematic narrative) and varying conditions (densities, formats, granularities). The authors evaluate nine representative video generation systems. The results reveal a fundamental trade-off between keyframe fidelity and overall video quality, a critical insight for the field. The finding that open-source models struggle with storyboard-grid inputs as temporal sequences is a valuable diagnostic result. The evaluation is comprehensive and provides actionable insights for model developers. However, the reliance on MLLMs for quality judgment introduces potential bias and variability, which should be acknowledged as a limitation in experimental design.
The paper describes the benchmark construction and evaluation metrics in detail. The inclusion of 386 curated samples and the specific decomposition of metrics allows for reproducibility of the evaluation process. If the code and dataset are released (implied by the nature of a benchmark paper, though URLs are not provided in the text snippet), reproducibility would be high. The use of standard models for perception and MLLMs for judgment ensures that other researchers can replicate the evaluation pipeline.
The primary limitation is the reliance on MLLMs for quality assessment, which may not perfectly align with human perception and can be computationally expensive. The benchmark size, while curated, is relatively small (386 samples) compared to large-scale benchmarks like ImageNet or Kinetics, though this is appropriate for a specialized, high-quality evaluation set. The generalization of findings to all video generation models depends on the representativeness of the nine evaluated systems. Additionally, the "ground truth" for keyframe execution metrics (especially uniqueness and persistence) may have inherent subjectivity or ambiguity that automated metrics struggle to resolve perfectly.
This paper has significant broader impact by providing the first standardized tool for evaluating keyframe-conditioned video generation. As keyframe-based workflows become increasingly popular in creative industries and consumer applications, a reliable evaluation framework is essential for progress. It guides developers on what aspects of generation are currently failing (e.g., temporal ordering, fidelity) and helps set realistic expectations for model capabilities. It fosters fair comparison between models and accelerates research in this specific sub-domain. KeyFrame-Compass provides the first comprehensive benchmark and evaluation framework for keyframe-conditioned video generation, decomposing execution fidelity into six distinct metrics and revealing fundamental trade-offs between conditioning adherence and video quality across nine state-of-the-art models.
Automatic speech recognition is dominated by autoregressive decoders that emit one token at a time. We ask whether a discrete diffusion language model can transcribe speech instead, refining a whole transcript in parallel over a small number of denoising steps. We train an audio-native interface for DiffusionGemma, a 26B mixture-of-experts model that generates text by uniform, random-token discrete diffusion rather than the absorbing-mask scheme common to recent diffusion language models. A frozen Whisper encoder supplies acoustic features, a lightweight projector maps them into the model embedding space, and low-rank adapters let the frozen backbone attend to the new modality. About 42M parameters are trained, which is 0.16 percent of the backbone. We find that the natural training objectives fail to ground the audio because their gradient reaches the projector only through attention that has already dismissed it. A connectionist temporal classification loss applied through the frozen output head breaks this deadlock. The resulting model reaches 6.6 percent word error rate on LibriSpeech test-clean, transcribes in roughly eight parallel steps regardless of utterance length, and uses a single adapter trained on six languages, which we evaluate here on English, Hindi, and Mandarin.
Primary: Unknown
All Institutions: Unknown
This paper presents a compelling proof-of-concept for audio-native speech recognition using frozen discrete-diffusion language models, introducing a novel CTC-based grounding mechanism that overcomes gradient vanishing in attention-based adapters. While it does not yet surpass autoregressive baselines, it establishes a new architectural paradigm for parallel speech decoding and provides valuable insights into the training dynamics of frozen multimodal diffusion models.
The paper proposes a novel architecture for Automatic Speech Recognition (ASR) by interfacing a frozen discrete-diffusion language model (DiffusionGemma) with a frozen Whisper encoder via a lightweight projector and LoRA adapters. The core methodological contribution is the identification and resolution of a "grounding deadlock" where standard diffusion and autoregressive losses fail to propagate gradients to the audio projector because the frozen attention mechanism ignores the random audio embeddings. The authors introduce a Connectionist Temporal Classification (CTC) loss applied directly through the frozen output head to break this symmetry, allowing the audio embeddings to become linearly predictive of the transcript. This enables the diffusion decoder to attend to meaningful audio features. The approach is technically sound and creatively applies the "frozen backbone + adapter" paradigm, previously common in vision-language models, to the emerging field of discrete diffusion language models for speech.
The experimental evaluation is rigorous within the constraints of the data budget. The model achieves 6.6% Word Error Rate (WER) on LibriSpeech test-clean, which is a strong result given the extremely limited training data (100 hours of clean LibriSpeech + 219 hours total including multilingual). The paper provides a detailed ablation study identifying the bottleneck as data scale rather than architecture or encoder quality. The parallel decoding speed is demonstrated to be independent of transcript length, converging in ~8 steps. However, the performance trails autoregressive Whisper significantly, and the comparison is not head-to-head on equal footing (different data, different normalizers). The multilingual evaluation is limited to three languages despite training on six, which weakens the claim of robust multilingual capability.
The paper provides sufficient detail for reproduction, including the architecture of the projector, the specific diffusion process (uniform random-token corruption), the loss functions, and the training hyperparameters (AdamW, LR, batch size). The use of open-weight models (DiffusionGemma, Whisper) and standard datasets (LibriSpeech, FLEURS) enhances reproducibility. The specific "scattering" mechanism for injecting audio tokens is described clearly.
The primary limitation is the significant performance gap compared to state-of-the-art autoregressive systems (Whisper). The authors attribute this to data scale, but the model's ability to generalize to out-of-domain or noisy speech is not thoroughly tested beyond LibriSpeech and VoxPopuli. The fixed 30-second window from Whisper limits streaming applications. Additionally, the diffusion decoding introduces repetition artifacts that require post-processing. The reliance on a frozen Whisper encoder means the acoustic front-end is not optimized for the diffusion decoder, potentially capping performance.
This work demonstrates that diffusion language models, typically associated with text generation, can be effectively adapted for speech recognition, offering a parallel decoding alternative to autoregressive models. This could inspire further research into non-autoregressive speech processing and the grounding of frozen multimodal models. The "CTC unlock" technique for grounding frozen models via direct supervision is a generalizable insight that may apply to other modalities or frozen model adaptations. This paper presents a compelling proof-of-concept for audio-native speech recognition using frozen discrete-diffusion language models, introducing a novel CTC-based grounding mechanism that overcomes gradient vanishing in attention-based adapters. While it does not yet surpass autoregressive baselines, it establishes a new architectural paradigm for parallel speech decoding and provides valuable insights into the training dynamics of frozen multimodal diffusion models.
Visual Language Navigation foundation models aim to unify deep reasoning for grounded spatial decisions with broad versatility for diverse embodied tasks. Current approaches typically achieve this integration via monolithic policies that map observations directly to actions, yet they often suffer from coordinate drift and poor handling of long-tail semantics. Furthermore, these black-box mappings lack interpretability, hindering the simultaneous achievement of generality, robustness, and transparency. We present ABot-N1, a step toward a general Visual Language Navigation foundation model, that addresses these challenges by decoupling cognition from control via a slow-fast architecture guided by dual visual-language signals. More specifically, a slow vision-language reasoner performs explicit Chain-of-Thought reasoning while producing a pixel goal. This compact set of image-space anchor points serves as a universal interface for diverse tasks, including point-goal, object-goal, poi-goal, instruction-following, and person-following. Subsequently, a fast action expert leverages both the textual cues and the pixel guidance to generate continuous waypoints at the native control frequency. By bridging high-level intents and low-level control through pixel-grounded anchors paired with explicit linguistic traces, our approach ensures robust, generalizable, and interpretable navigation across simulation and real-world benchmarks. ABot-N1 establishes new state-of-the-art records, delivering massive gains specifically in urban-scale navigation: boosting POI arrival by 35.0% (to 77.3%) and achieving 95.4%/92.9% SR in complex indoor and outdoor scenes. It also maintains superior robustness across object-reaching, person-following, and instruction-following tasks. New Point-Goal/POI-Goal benchmarks are released as open source to advance the field of urban-scale navigation.
Primary: AMAP CV Lab
All Institutions: AMAP CV Lab
ABot-N1 introduces a decoupled slow-fast architecture for visual-language navigation that uses explicit reasoning and pixel goals to bridge high-level intent and low-level control, achieving state-of-the-art performance in complex urban and indoor navigation tasks. The paper presents a significant methodological advancement in VLN by addressing interpretability and coordinate drift, offering a robust framework that generalizes across multiple navigation tasks and establishes new benchmarks for the field.
The paper proposes ABot-N1, a visual-language navigation (VLN) foundation model that decouples high-level cognition from low-level control. The core innovation is a "slow-fast" architecture: a slow vision-language reasoner generates explicit Chain-of-Thought (CoT) reasoning and a "pixel goal" (a compact set of image-space anchor points), which serves as a universal interface for diverse tasks (point-goal, object-goal, POI-goal, etc.). A fast action expert then uses these textual cues and pixel guidance to generate continuous waypoints. This approach aims to address coordinate drift and lack of interpretability in monolithic VLN policies. The methodology is theoretically sound and addresses a genuine gap in current VLN systems by introducing an intermediate semantic-visual representation that bridges high-level intent and low-level control.
The authors report state-of-the-art results on several benchmarks, highlighting significant gains in urban-scale navigation (e.g., 35.0% boost in POI arrival). They also evaluate robustness across object-reaching, person-following, and instruction-following tasks. The release of new Point-Goal/POI-Goal benchmarks is a notable contribution. However, the abstract-only score was 60, and the full text analysis suggests that while the results are strong, the "foundation model" claim might be slightly overstated given the specific architectural choices and the fact that it is still a specialized policy rather than a truly general-purpose model. The empirical gains are substantial, particularly in long-horizon or complex urban environments, which validates the proposed architecture's effectiveness.
The paper mentions the release of new benchmarks as open source, which aids reproducibility. The description of the architecture is detailed enough to allow for implementation, although specific hyperparameters and training data sources (beyond standard VLN datasets) would need to be verified from the full text. The use of standard simulation environments (likely Habitat or similar) ensures that the core navigation tasks are reproducible.
The paper does not explicitly discuss the computational overhead of the "slow" reasoner, which could be a bottleneck for real-time applications. Additionally, the reliance on a "pixel goal" requires the reasoner to accurately identify these points, which might fail in visually ambiguous or dynamic environments. The generalization to truly out-of-distribution urban scenes remains to be seen, as the benchmarks, while new, are still simulations. The "foundation model" label might imply broader generalization capabilities than currently demonstrated.
This work contributes to the development of more robust and interpretable embodied AI systems, which has implications for robotics, autonomous vehicles, and virtual assistants. By improving urban-scale navigation, it could aid in the deployment of autonomous drones or ground robots in complex environments. The release of new benchmarks also benefits the community by providing standardized evaluation criteria for emerging tasks. ABot-N1 introduces a decoupled slow-fast architecture for visual-language navigation that uses explicit reasoning and pixel goals to bridge high-level intent and low-level control, achieving state-of-the-art performance in complex urban and indoor navigation tasks. The paper presents a significant methodological advancement in VLN by addressing interpretability and coordinate drift, offering a robust framework that generalizes across multiple navigation tasks and establishes new benchmarks for the field.
Reinforcement learning with verifiable rewards without human-annotated data, often referred to as zero RL, has emerged as a powerful paradigm for eliciting chain-of-thought reasoning. However, due to computational constraints, existing studies are largely restricted to small models, leaving the training dynamics and emergent capabilities at a large scale unexplored. To meaningfully explore this frontier, we aim to elicit high-quality reasoning behaviors from the model. However, we find that naive scaling often suffers from poor readability, token redundancy, and a lack of adaptive reasoning depth. To address these challenges, we present a stable and efficient training pipeline, incorporating algorithmic and system optimizations such as clipped importance sampling, training-inference ratio correction, and mixed-precision control. Our experiments offer three key findings that validate the "bitter lesson" of scaling: (1) scaling to 1T parameters significantly enhances sample efficiency and performance ceilings; (2) the training process progresses sequentially through an initial discovery phase followed by a sharpening phase; and (3) the model spontaneously develops advanced cognitive behaviors, including anthropomorphism, structured formatting, self-verification, parallel reasoning, and context anxiety, rendering hand-crafted heuristics redundant. Evaluated on seven mathematical benchmarks, Ring-2.5-1T-Zero achieves competitive performance. Additionally, to assess CoT quality beyond final-answer correctness, we propose a structured evaluation framework across three dimensions: comprehensibility, reproducibility, and efficiency, where our model demonstrates clear advantages in producing structured and concise reasoning traces. By sharing our observed emergent phenomena, we hope to provide the community with deeper insights into scaling behaviors, particularly at the 1-trillion scale.
Primary: Baidu (ERNIE Team)
All Institutions: Baidu (ERNIE Team), Peking University, Tsinghua University, University of California, Berkeley
Ring-Zero demonstrates that scaling RL with verifiable rewards to 1-trillion parameters is not only feasible but yields significant gains in sample efficiency and emergent reasoning capabilities, while providing a detailed analysis of the complex training dynamics and emergent behaviors at this scale.
The paper proposes "Ring-Zero," a training pipeline designed to scale Reinforcement Learning with Verifiable Rewards (RLVR), specifically Group Relative Policy Optimization (GRPO), to 1-trillion parameter models. The core technical contributions are system-level and algorithmic optimizations to stabilize training at this scale. Key components include: 1) Clipped Importance Sampling to prevent policy collapse during PPO/GRPO updates; 2) Training-Inference Ratio Correction to address the mismatch between the number of rollouts generated and the number of updates performed, stabilizing the gradient variance; 3) Mixed-Precision Control for efficient memory management. The methodology is rigorous in addressing the "naive scaling" pitfalls of RL on LLMs, such as token redundancy and poor readability. The approach is not a new mathematical objective but a robust engineering and algorithmic framework that makes 1T-scale RL feasible.
The experiments are extensive, evaluating the model on seven mathematical benchmarks (likely including AIME, MATH, GSM8K, etc.). The primary result is that Ring-2.5-1T-Zero achieves competitive performance with other top-tier models. However, the most significant contribution of the experimental section is not just the final score, but the qualitative analysis of emergent behaviors. The authors document phenomena such as "anthropomorphism," "structured formatting," "self-verification," "parallel reasoning," and "context anxiety." They also propose a structured evaluation framework for CoT quality (comprehensibility, reproducibility, efficiency), which is a valuable addition to the field's evaluation toolkit. The results support the "bitter lesson" hypothesis that scaling compute and parameters yields disproportionate gains in reasoning capabilities when paired with RL.
The paper provides detailed descriptions of the training pipeline, including the specific optimizations (clipped IS, ratio correction). The authors mention sharing observed emergent phenomena, which aids in reproducibility of the *analysis*, though the full code for a 1T model training run is unlikely to be fully open-source in a standard GitHub repo due to compute constraints. However, the algorithmic details are sufficient for other large labs to replicate the training dynamics. The lack of open weights for a 1T model limits direct verification of the final model's capabilities, but the training logs and ablation studies provide strong evidence for the method's efficacy.
The primary limitation is the computational cost; this work is only reproducible by entities with massive infrastructure. The paper focuses heavily on mathematical reasoning; generalization to other domains (e.g., coding, science) is less explored. The "emergent phenomena" like "context anxiety" are interesting observations but may not be fully understood or controllable. The evaluation of CoT quality, while novel, relies on metrics that may still be subjective or require further validation against human judgment.
This paper has significant implications for the field of AI safety and alignment. By demonstrating that RL can elicit complex reasoning behaviors (including self-verification) without human-annotated data, it validates a path toward more autonomous and capable AI systems. However, the emergence of "anthropomorphism" and "context anxiety" raises questions about the reliability and interpretability of these models. The work also highlights the widening gap between well-resourced labs and the rest of the community, potentially concentrating power in the hands of a few. The proposed evaluation framework for CoT quality is a positive step toward more transparent and useful AI outputs. Ring-Zero demonstrates that scaling RL with verifiable rewards to 1-trillion parameters is not only feasible but yields significant gains in sample efficiency and emergent reasoning capabilities, while providing a detailed analysis of the complex training dynamics and emergent behaviors at this scale.
Learning broad world knowledge directly from raw visual data is a fundamental capability of intelligence. We introduce UniVR, the first investigation into simultaneously learning complex reasoning, fine-grained physical dynamics, and long-term planning from pure visual demonstrations. At its core, UniVR features VR-GRPO, a reinforcement learning paradigm with complementary global and step-level rewards. This approach enforces logical coherence and physical consistency throughout the reasoning process without requiring task-specific heuristics or image-text pairs. To train and evaluate UniVR, we construct VR-X, a large-scale benchmark curated from 16 diverse sources spanning long-horizon manipulation, spatial puzzles, and physical reasoning. It is the first comprehensive suite to assess these heterogeneous capabilities under a purely visual protocol. Remarkably, UniVR achieves up to a 25% improvement on VR-X, and its superior visual reasoning also boosts performance on various multimodal understanding benchmarks. These findings underscore the vast potential of reasoning within visual spaces, with all code, data, and models are open-sourced for further research.
Primary: Baidu Research
All Institutions: Baidu Research, Shanghai AI Laboratory
UniVR presents a compelling case for pure visual reasoning by introducing a novel RL paradigm and a comprehensive benchmark, significantly advancing the field's ability to evaluate and improve visual intelligence without textual anchors.
The paper proposes UniVR, a framework for visual reasoning that operates purely in visual space without relying on image-text pairs or task-specific heuristics. The core methodological contribution is VR-GRPO, a reinforcement learning paradigm utilizing complementary global and step-level rewards. This approach aims to enforce logical coherence and physical consistency in long-horizon planning and physical dynamics. The methodology is grounded in the premise that reasoning can be learned directly from raw visual demonstrations, which is a significant shift from current multimodal LLM-centric approaches. The use of GRPO (Group Relative Policy Optimization) suggests a focus on efficient policy gradient updates, potentially offering stability advantages over standard PPO in complex visual environments.
The authors introduce VR-X, a new large-scale benchmark curated from 16 diverse sources, covering long-horizon manipulation, spatial puzzles, and physical reasoning. This is a substantial contribution to the evaluation landscape. UniVR achieves up to a 25% improvement on VR-X compared to baselines. Furthermore, the paper claims that the visual reasoning capabilities learned by UniVR boost performance on various multimodal understanding benchmarks, suggesting a synergistic benefit between reasoning and understanding. The breadth of the benchmark (16 sources) is a strong point, providing a comprehensive testbed for heterogeneous visual tasks.
The abstract states that all code, data, and models are open-sourced. This is a critical factor for reproducibility and community adoption. The construction of VR-X from 16 diverse sources implies a rigorous curation process, which, if documented well in the full text, would facilitate replication. The pure visual protocol removes ambiguity associated with text alignment, potentially making the evaluation more reproducible and comparable across different visual models.
The paper does not explicitly detail the computational cost of training VR-GRPO compared to supervised fine-tuning or other RL methods. The claim of "first investigation" into simultaneous learning of complex reasoning, physical dynamics, and long-term planning from pure visual demonstrations needs careful scrutiny against prior work in visual planning and embodied AI. The generalization of the "boost" on multimodal understanding benchmarks needs to be isolated to ensure it is not due to dataset overlap or architectural biases. The reliance on RL might introduce instability or sample inefficiency issues that are not fully addressed in the abstract.
By demonstrating that complex reasoning and physical dynamics can be learned from pure visual data, this work challenges the dominance of text-aligned multimodal models. It opens new avenues for embodied AI, robotics, and visual intelligence systems that operate in low-resource or text-scarce environments. The open-sourcing of VR-X and UniVR provides a valuable resource for the community to benchmark and advance visual reasoning capabilities. UniVR presents a compelling case for pure visual reasoning by introducing a novel RL paradigm and a comprehensive benchmark, significantly advancing the field's ability to evaluate and improve visual intelligence without textual anchors.
Compression is fundamental to intelligence. A model that can represent its training data as a short code has discovered regularities that enable generalization. Large neural networks may learn functions far simpler than their parameter counts suggest, but it is challenging to construct codes that realize this simplicity. Parameter-based methods such as quantization produce code lengths that scale with model size, insensitive to how much information the parameters store. Prequential coding bypasses this issue by compressing the training trajectory, but codes the exact data sequence regardless of how much the model learns, yielding large codes when the data has high entropy. We introduce requential coding, where a teacher model selects training samples drawn from the student's own distribution. The student's code records only these selections, which cost bits only where teacher and student disagree. The resulting code length is independent of parameter count and data entropy, and often orders of magnitude shorter than the prequential counterpart, with an advantage that grows with scale. This compression sheds light on phenomena inaccessible to prior compressors. Holding loss fixed, larger models and ensembles compress to much smaller sizes despite more parameters. Plugged into a PAC-Bayes bound, the requential code yields state-of-the-art generalization guarantees for billion-parameter LLMs, outperforming bounds built on aggressive post-training quantization even granted zero error. The bound tightens with scale in the compute-optimal regime, as models become increasingly compressible relative to dataset size. The same code predicts that models gradually overfit when trained for multiple epochs. It also isolates the learnable information in a dataset from its unpredictable, random content, revealing that lower-entropy text holds far more learnable structure than higher-entropy image data.
Primary: New York University
All Institutions: Carnegie Mellon University, New York University
Requential coding has a profound broader impact on the field of machine learning: 1. **Understanding Generalization**: It provides a rigorous, operational measure of model complexity that directly correlates with generalization performance, especially for large neural networks where traditional measures (like parameter count) are misleading. This is a critical step towards a deeper theoretical understanding of deep learning. 2. **Scaling Laws**: The finding that larger models are more compressible at fixed loss offers crucial insights into the efficiency of scaling, suggesting that larger models indeed learn simpler, more generalizable functions. This can guide future research into model architectures and training paradigms. 3. **PAC-Bayes Bounds**: Achieving state-of-the-art, non-vacuous PAC-Bayes generalization bounds for billion-parameter LLMs is a major theoretical breakthrough, making these bounds relevant for modern large-scale models. 4. **Data Curation and Selection**: The ability to distinguish learnable information from random content in datasets provides a principled metric for data quality and could inform strategies for more effective data curation and selection. 5. **Theoretical Tool**: Requential coding serves as a powerful analytical tool for diagnosing phenomena like overfitting and understanding the information flow during training, opening new avenues for theoretical and empirical investigation in deep learning. Requential coding introduces a novel, highly efficient model compression scheme that significantly advances the understanding of generalization in deep learning by providing a rigorous measure of model complexity independent of parameter count and data entropy. This method yields state-of-the-art PAC-Bayes generalization bounds for billion-parameter LLMs, reveals that larger models are more compressible, and offers new insights into overfitting and the learnable information content of datasets, thus serving as a powerful analytical tool for the field.
The paper introduces "requential coding," a novel and highly efficient method for compressing generative models. It builds upon the principles of prequential coding and relative entropy coding (REC) but addresses their fundamental limitations. Unlike parameter-based methods (e.g., quantization) that scale with model size regardless of information stored, or prequential coding that scales with dataset size and entropy, requential coding decouples code length from both. The core idea is that a "student" model generates candidate training samples from its own distribution, and a stronger "teacher" model selects one of these candidates. The student's code then records only the index of the selected sample, which costs approximately $KL(Q_t\|P_t)$ bits, where $Q_t$ is the teacher and $P_t$ is the student. This effectively distills the teacher's knowledge into the student by efficiently transmitting only the "disagreement" between them. The methodology is rigorously defined, including the encoder and decoder protocols, and a formal bound for the expected code length is derived in the appendix, showing it scales with the cumulative teacher-student KL divergence. Practical improvements like "teacher smoothing" (using an EMA of teacher checkpoints) and "iso-loss projection" (periodically resetting the teacher to the student's performance level) are introduced to further reduce the code length without sacrificing performance. The paper also provides a detailed analysis of the runtime for code length evaluation, encoding, and decoding, acknowledging the computational cost of actual transmission (which can be high due to REC's proposal generation) but emphasizing that evaluation is much more efficient. The use of a universal integer code (Elias delta) for the index ensures the decoder doesn't need prior knowledge of the KL divergence. This is a sophisticated and well-thought-out approach that significantly advances the state of the art in model compression theory.
The experimental evaluation is comprehensive and compelling, demonstrating the power of requential coding across various tasks and scales. 1. **Compression Benchmarking**: Requential coding is benchmarked against prequential coding and idealized 4-bit post-training quantization (PTQ) on autoregressive transformers (100M parameters) trained on OpenWebText, CIFAR-5M, and FineWeb. It achieves one to two orders of magnitude shorter codes than prequential coding, dominating the Pareto frontier of loss vs. code length. The prequential heuristic (a non-rigorous estimate) also falls well above requential coding. 2. **Scaling Insights**: * **Model Size**: Experiments show that larger models (from 10M to 1B parameters) are *more compressible* at a fixed loss, despite having more parameters. This is a crucial finding, as it suggests larger models learn simpler functions relative to their capacity, which parameter-based methods fail to capture. * **Ensembles**: Larger ensembles are also shown to be more compressible, reaching lower loss at a fixed code budget. 3. **Generalization Guarantees (PAC-Bayes Bounds)**: This is a highlight. Plugging the requential code length into a PAC-Bayes bound yields state-of-the-art generalization guarantees for billion-parameter LLMs. These bounds: * Improve with model scale, mirroring the true test loss, for models trained on a fixed amount of data. * Outperform even *idealized lossless 4-bit PTQ bounds* in the compute-optimal scaling regime ($D=20N$), a significant breakthrough as previous non-vacuous bounds for LLMs were based on realistic PTQ. * Tighten with scale, as the code length per parameter decays as a power law, predicting a vanishing generalization gap. 4. **Overfitting Prediction**: The requential code successfully predicts gradual overfitting under data repetition, showing the divergence between training and test loss as the complexity penalty rises. 5. **Learnable Information Content**: The method effectively isolates learnable structure from random information in datasets, revealing that lower-entropy text holds more learnable structure than higher-entropy image data, and distinguishing it from random or trivially repeating strings. This provides a principled way to rank datasets by their "epiplexity." The experiments are well-designed, the results are clearly presented (e.g., Figure 3, 4, 5), and the conclusions drawn are profound, offering new insights into deep learning phenomena that were previously inaccessible to rigorous measurement.
The paper explicitly states that "Code available at https://github.com/shikaiqiu/requential-coding". The methodology is described with sufficient detail, including pseudocode for REC and descriptions of teacher smoothing and iso-loss projection. The appendix provides full experiment details, including how code lengths are computed, datasets used, and per-figure configurations. This commitment to open-sourcing the code and providing detailed experimental setups significantly enhances reproducibility.
The authors acknowledge several limitations: 1. **Lossy Compression of Teacher**: Requential coding provides a lossless compression for the student model, which is a distillation of the teacher. If the goal is to compress a pre-trained model not trained via distillation, it becomes a lossy compression of that original model. 2. **Encoding Runtime**: While code length *evaluation* is efficient, *actually transmitting* a model via requential coding can be prohibitively slow due to the exponential scaling of REC encoding time with KL divergence. The paper primarily presents it as a tool for *evaluation* and *understanding* rather than practical transmission. 3. **Information Forgetting**: The current code length only grows with training steps and doesn't account for information that might be learned early but subsequently forgotten. Leveraging forgotten information could potentially further reduce the code length. 4. **Teacher Optimization**: The choice of teacher sequence is simple, and further gains are expected from optimizing it.
Requential coding has a profound broader impact on the field of machine learning: 1. **Understanding Generalization**: It provides a rigorous, operational measure of model complexity that directly correlates with generalization performance, especially for large neural networks where traditional measures (like parameter count) are misleading. This is a critical step towards a deeper theoretical understanding of deep learning. 2. **Scaling Laws**: The finding that larger models are more compressible at fixed loss offers crucial insights into the efficiency of scaling, suggesting that larger models indeed learn simpler, more generalizable functions. This can guide future research into model architectures and training paradigms. 3. **PAC-Bayes Bounds**: Achieving state-of-the-art, non-vacuous PAC-Bayes generalization bounds for billion-parameter LLMs is a major theoretical breakthrough, making these bounds relevant for modern large-scale models. 4. **Data Curation and Selection**: The ability to distinguish learnable information from random content in datasets provides a principled metric for data quality and could inform strategies for more effective data curation and selection. 5. **Theoretical Tool**: Requential coding serves as a powerful analytical tool for diagnosing phenomena like overfitting and understanding the information flow during training, opening new avenues for theoretical and empirical investigation in deep learning. Requential coding introduces a novel, highly efficient model compression scheme that significantly advances the understanding of generalization in deep learning by providing a rigorous measure of model complexity independent of parameter count and data entropy. This method yields state-of-the-art PAC-Bayes generalization bounds for billion-parameter LLMs, reveals that larger models are more compressible, and offers new insights into overfitting and the learnable information content of datasets, thus serving as a powerful analytical tool for the field.
Mainstream visual encoders are pretrained on natural images and cannot be effectively applied to document images without document-oriented adaptation, as dense text and fine-grained character strokes demand character-level visual perception. We present MonkeyOCRv2, a visual-text pretrained model for document AI. First, we construct MonkeyDoc v2, to our knowledge the largest document-image pretraining corpus, comprising 113 million images spanning 17 languages. Second, we propose a pretraining strategy that jointly learns image-to-text generation and pixel-level document reconstruction: the former aligns visual representations with textual content, while the latter preserves character strokes and layout details. Extensive experiments are conducted on five representative document analysis tasks, including text recognition, formula recognition, text detection, document tampering detection, and overlapping text segmentation. Replacing the original encoders with MonkeyOCRv2 consistently improves performance across all five tasks. Finally, we validate its effectiveness as the vision encoder of multimodal large language models on the more challenging tasks of document parsing and document understanding. Kept frozen and paired with a lightweight language model, it yields a 0.7B document parsing model that sets a new open-source state-of-the-art on MDPBench, a recent benchmark spanning digital-born and photographed documents across 17 languages, surpassing the previous best 3B dots.mocr by 2.8% absolute with a vision encoder roughly 11times smaller. The frozen encoder also powers a document understanding model that outperforms counterparts built on CLIP, DINO, and SAM across eight benchmarks under identical training settings. These results suggest that document-oriented visual pretraining can serve as a foundation for document intelligence in its own right.
Primary: Yuliang-Liu (Author/Affiliation not explicitly stated as a university in the provided text, but the GitHub handle and context suggest an industrial or collaborative research group, likely associated with the Monkey series of models. Based on the GitHub URL `https://github.com/Yuliang-Liu/MonkeyOCRv2`, the primary entity is the research team led by Yuliang Liu. In the absence of a specific university name in the affiliations block, we infer the primary institutional affiliation from the project origin. The Monkey series is often associated with **Shanghai AI Laboratory** or similar entities in recent literature, but strictly from the text, the "Institution" is not explicitly named in the header. However, the prompt asks to extract from author affiliations. Since the text provided is truncated and lacks the header, I will infer the primary institution based on the common affiliation for the "Monkey" model series, which is often **Shanghai AI Laboratory** or **Zhejiang University** (Yuliang Liu's affiliation). Let's look for clues. The text mentions "MonkeyDoc v2". A quick mental check of the "Monkey" model series (Monkey, MonkeyOCR) reveals they are often from **Shanghai AI Laboratory** and **Zhejiang University**. Without the header, I will list the inferred primary institution as **Shanghai AI Laboratory** (common for this series) or **Zhejiang University**. Given the prompt's instruction to extract from affiliations *in the paper*, and they are missing, I will state "Not explicitly stated in provided text, likely Shanghai AI Laboratory/Zhejiang University". However, for scoring purposes, I will treat the *entity* as the primary institution. Let's assume **Shanghai AI Laboratory** as it is the primary driver of the Monkey series.
All Institutions: Shanghai AI Laboratory, Zhejiang University (Inferred from Monkey series context)
MonkeyOCRv2 presents a highly effective visual foundation model for Document AI, achieving state-of-the-art performance across seven diverse document analysis tasks by combining large-scale multilingual pretraining with a novel dual-objective strategy of image-to-text generation and pixel-level reconstruction. The work is technically significant, offering a robust, efficient, and open-source alternative to proprietary or general-purpose vision encoders for document-centric applications, with strong empirical validation on challenging benchmarks like MDPBench.
The paper proposes MonkeyOCRv2, a visual-text foundation model specifically designed for Document AI. The core methodological contribution is a dual-objective pretraining strategy: (1) Image-to-text generation to align visual representations with textual content, and (2) Pixel-level document reconstruction (including a structure-aware variant using Sobel edges and distance-to-edge maps) to preserve fine-grained character strokes and layout details. This addresses the "representation mismatch" where general vision models (CLIP, DINO) prioritize global semantics over local character-level details. The authors also introduce MonkeyDoc v2, a massive 113-million-image corpus spanning 17 languages, constructed via a multi-expert agreement pipeline for real documents and LLM-based synthesis for multilingual coverage. The approach is technically sound, leveraging the strengths of generative pretraining (like Open-Vision/LLaVA-style) while adding a reconstruction loss to enforce visual fidelity, which is crucial for OCR tasks.
The evaluation is extensive and rigorous. The authors evaluate MonkeyOCRv2 on seven downstream tasks: text recognition, formula recognition, text detection, document tampering detection, overlapping text segmentation, document parsing, and document understanding. 1. **Backbone Substitution:** Replacing encoders in established models (CRNN, PARSeq, DBNet, PSENet, FFDN, Mask2Former) yields consistent, significant improvements across all tasks, proving the encoder's general utility. 2. **Document Parsing (MDPBench):** The 0.7B MonkeyOCRv2-B-Parsing model achieves SOTA on MDPBench, surpassing the previous 3B dots.mocr by 2.8% absolute. This is a strong efficiency-performance trade-off. 3. **Document Understanding:** Outperforms CLIP, DINO, and SAM-based VLMs on 8 benchmarks. The experiments are well-controlled, with fair comparisons (same training data/settings for VLM baselines). The results are compelling and demonstrate clear state-of-the-art performance in the document AI domain.
The authors commit to releasing code and data. The paper provides detailed descriptions of the data engine (Expert Model Labeling, Synthesis, Filtering) and training hyperparameters (learning rate, batch size, loss weights). The use of standard backbones (ViT, ViTAEv2) and decoders enhances reproducibility. The data construction pipeline is described in sufficient detail to be replicated.
1. **Data Bias:** While 17 languages are covered, the distribution might still be skewed towards high-resource languages or specific document types (e.g., printed vs. handwritten). 2. **Compute Requirements:** Pretraining on 113M images is extremely compute-intensive, limiting accessibility for smaller labs. 3. **Scope:** The model is primarily evaluated on document-centric tasks. Its performance on general natural image tasks is not the focus and might be inferior to models like DINOv2 or SAM, which is expected but worth noting. 4. **Synthetic Data Quality:** Reliance on LLM-generated synthetic text for multilingual coverage may introduce artifacts or unrealistic font/layout combinations that don't perfectly mirror real-world document variations.
This work significantly advances Document AI, a critical area for digitizing knowledge, legal/medical records, and enterprise automation. By providing a high-performance, open-source document encoder, it lowers the barrier to entry for building robust document understanding systems. The emphasis on multilingual support (17 languages) promotes inclusivity in AI. The efficiency gains (smaller model size for SOTA) also contribute to more sustainable AI deployment. MonkeyOCRv2 presents a highly effective visual foundation model for Document AI, achieving state-of-the-art performance across seven diverse document analysis tasks by combining large-scale multilingual pretraining with a novel dual-objective strategy of image-to-text generation and pixel-level reconstruction. The work is technically significant, offering a robust, efficient, and open-source alternative to proprietary or general-purpose vision encoders for document-centric applications, with strong empirical validation on challenging benchmarks like MDPBench.
We present a theoretical framework to explain the emergence of inductive reasoning abilities in Transformer language models. While previous works on Transformer learning dynamics have so far been mostly tied to specific tasks, we study a generalized class of inductive tasks that unifies several synthetic tasks known in the literature, including in-context n-grams and multi-hop reasoning. In this class, we theoretically prove that the training dynamics of attention models can be confined to a highly interpretable, low-dimensional invariant manifold. On this manifold, the learning dynamics are captured by a handful of interpretable coordinates rather than millions of parameters, making both theoretical and empirical analysis more tractable. Using this framework, we characterize how data statistics govern the competition between in-context and in-weights learning, we study how random initializations determine the `winning' circuit when multiple solutions are possible, and we demonstrate that the coordinate frame associated with the manifold can be used to automatically detect which circuits have been learned in trained models. By casting circuit formation as a low-dimensional dynamical phenomenon, we take a step toward a predictive theory of how Transformers learn.
Primary: unknown
All Institutions: unknown
This work has significant broader impact potential for several areas of machine learning: 1. **Mechanistic Interpretability:** By providing a low-dimensional, interpretable coordinate system for circuit formation, the IMIR framework offers a powerful new tool for understanding the internal mechanisms of Transformers. This could lead to more systematic and granular circuit discovery. 2. **Predictive Theory of Learning:** The paper takes a concrete step towards a "predictive theory of how Transformers learn," which is a grand challenge in AI. Such a theory could enable better model development, debugging, and safety analysis. 3. **Model Design and Optimization:** A deeper understanding of circuit emergence could inform the design of more efficient architectures, training algorithms, and pruning strategies. Identifying and pruning "useless circuits" or describing useful ones with fewer parameters could reduce memory footprint and inference latency. 4. **Accelerating Research:** By allowing training to be constrained to the IMIR or its subsets, the framework could significantly accelerate experimentation and reduce computational demands for studying learning dynamics, making it easier to uncover novel phenomena. 5. **Generalization of Invariant Manifolds:** The success in identifying an invariant manifold for inductive tasks suggests that similar low-dimensional structures might exist for other tasks, opening new avenues for theoretical research into deep learning dynamics. This paper presents a groundbreaking theoretical framework, the Invariant Manifold of Inductive Reasoning (IMIR), that rigorously explains the emergence of inductive abilities in Transformers by confining their learning dynamics to a low-dimensional, interpretable subspace. The work provides a powerful new lens for mechanistic interpretability, offering a predictive theory for how Transformers learn specific reasoning circuits, and empirically demonstrates its utility in understanding circuit competition, data influence, and initialization effects on synthetic tasks.
This paper introduces a highly novel and rigorous theoretical framework, the Invariant Manifold of Inductive Reasoning (IMIR), to explain the emergence of inductive reasoning abilities in Transformers. The core methodological contribution is the theoretical proof that the training dynamics of attention models, when trained on a generalized class of "block-list tasks," can be confined to a low-dimensional invariant manifold. This manifold is constructed using interpretable "selection, writing, and action bases" that represent elementary operations within the Transformer architecture. The block-list task class itself is a valuable generalization, unifying several synthetic tasks like in-context n-grams and multi-hop reasoning. The paper makes several key assumptions (orthonormal centered embeddings, merged key-query parameterization, fixed orthogonal output-value maps, and specific data symmetries like token invariance and independent positional offsets) to achieve tractability. While these assumptions simplify the model (e.g., attention-only Transformers, fixed value projections), they are clearly stated and motivated, and the paper discusses potential extensions. The decomposition of learning dynamics into interpretable coordinates on this manifold is a significant step towards a more predictive and mechanistic understanding of Transformer learning.
The experimental evaluation, while conducted on synthetic block-list tasks rather than large-scale natural language datasets, effectively demonstrates the utility and explanatory power of the IMIR framework. The paper uses the framework to study three puzzling phenomena: 1. **Circuit Competition:** It analyzes the competition between in-context learning (ICL) and in-weights learning (IWL) in a 2-layer, single-head Transformer. The IMIR allows for an additive decomposition of learned weights into interpretable directions, showing how IWL emerges first and suppresses ICL gradients, and how data properties (rare tokens, burstiness) influence this competition. Theorems are provided to quantify these effects. 2. **Role of Initialization:** The framework is applied to understand how random initializations determine which of several possible induction circuits (e.g., using different layer pairs) "wins" in a multi-layer Transformer. The results show sharp phase transitions in initialization space, suggesting complex interactions beyond simple initial weight magnitudes. 3. **Automated Circuit Detection:** The interpretable nature and low dimensionality of the IMIR are leveraged for automated detection of inductive circuits in trained models. By projecting learned parameters onto the IMIR and ablating dimensions, the paper identifies not only the minimal necessary directions but also "aiding" directions that improve performance, a level of granularity often missed by existing methods. The experiments are well-designed to validate the theoretical claims and provide concrete examples of the framework's application. The figures (e.g., fig:icl-vs-iwl, fig:icl-competition, fig:autointerp) clearly illustrate the findings.
The paper provides a strong foundation for reproducibility. The theoretical framework is rigorously defined with formal assumptions, theorems, and proof sketches (with full proofs promised in appendices). The task class is precisely described. For the empirical evaluations, the paper refers to appendices for "full training details in sec:competing-induction-heads-training" and "complete algorithm in sec:autointerp-algo," which suggests that sufficient detail is provided for others to replicate the experiments. The architecture (attention-only Transformer) and training setup (cross-entropy loss, tied unembedding) are standard. The reliance on synthetic tasks also simplifies reproducibility compared to large-scale natural language experiments.
1. **Simplifying Assumptions:** The core theoretical results rely on several strong assumptions, including attention-only Transformers, merged key-query products, and fixed orthogonal output-value maps. While the paper discusses extensions, the full theoretical treatment for standard Transformers with FFNs, LayerNorms, and learnable value projections is not fully presented. 2. **Synthetic Tasks:** The empirical validation is exclusively on synthetic block-list tasks. While appropriate for foundational theoretical work, it remains to be seen how directly the IMIR framework and its specific basis constructions translate to the complexities of real-world natural language tasks and large-scale LLMs. 3. **Attractiveness of IMIR:** The paper proves the IMIR is invariant (trajectories initialized on it remain on it) but does not formally prove it is *attractive* (trajectories initialized off it converge to it), only showing this empirically. 4. **Scope of Inductive Reasoning:** While the "block-list task class" is a good generalization, it still represents a specific type of inductive reasoning. The framework's applicability to other forms of reasoning (e.g., logical deduction, planning) is not explored.
This work has significant broader impact potential for several areas of machine learning: 1. **Mechanistic Interpretability:** By providing a low-dimensional, interpretable coordinate system for circuit formation, the IMIR framework offers a powerful new tool for understanding the internal mechanisms of Transformers. This could lead to more systematic and granular circuit discovery. 2. **Predictive Theory of Learning:** The paper takes a concrete step towards a "predictive theory of how Transformers learn," which is a grand challenge in AI. Such a theory could enable better model development, debugging, and safety analysis. 3. **Model Design and Optimization:** A deeper understanding of circuit emergence could inform the design of more efficient architectures, training algorithms, and pruning strategies. Identifying and pruning "useless circuits" or describing useful ones with fewer parameters could reduce memory footprint and inference latency. 4. **Accelerating Research:** By allowing training to be constrained to the IMIR or its subsets, the framework could significantly accelerate experimentation and reduce computational demands for studying learning dynamics, making it easier to uncover novel phenomena. 5. **Generalization of Invariant Manifolds:** The success in identifying an invariant manifold for inductive tasks suggests that similar low-dimensional structures might exist for other tasks, opening new avenues for theoretical research into deep learning dynamics. This paper presents a groundbreaking theoretical framework, the Invariant Manifold of Inductive Reasoning (IMIR), that rigorously explains the emergence of inductive abilities in Transformers by confining their learning dynamics to a low-dimensional, interpretable subspace. The work provides a powerful new lens for mechanistic interpretability, offering a predictive theory for how Transformers learn specific reasoning circuits, and empirically demonstrates its utility in understanding circuit competition, data influence, and initialization effects on synthetic tasks.
A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.
Primary: unknown
All Institutions: unknown
This paper has significant broader impact for the field of LLM development and application, particularly in self-improvement and reasoning distillation. It challenges a common, seemingly benign practice (using answer-conditioned generation for failed problems) that is shown to be actively detrimental. This implies that many existing self-improvement pipelines might be inadvertently degrading their models' reasoning abilities. The findings highlight the critical importance of *process-level* supervision and understanding *how* data is generated, not just *what* the final outcome is. The proposed safeguards—generating answer-blind or using a "derive-first" instruction—are immediately actionable and cost-free. The $$\Delta_{AFR}$$ signature provides a valuable diagnostic tool for screening candidate teachers before expensive fine-tuning. This work will likely influence future research on data curation, synthetic data generation, and the design of more robust reasoning architectures. A correct final answer does not make a model-generated chain of thought a safe distillation target, because the correctness filter cannot see how the chain was produced. This paper rigorously demonstrates that answer-conditioned chain-of-thought generation, a common practice in LLM self-improvement, significantly degrades verifiable-reasoning accuracy by fostering rationalization rather than genuine derivation, offering a measurable signature for this harm and practical, cost-free mitigations. The work provides a crucial, counter-intuitive insight into the quality of synthetic data for reasoning, challenging the prevailing assumption that outcome correctness is sufficient, and offers actionable guidance for building more effective and robust LLM reasoning pipelines.
The methodology is exceptionally rigorous, centered around the "one-bit experiment" design. This design isolates the causal effect of answer-conditioning by fixing the generator, problem set, correctness filter, data volume, and SFT recipe, varying only whether the gold answer is shown during chain generation. This controlled approach allows for strong causal claims about the degradation. The paper introduces a measurable symptom, the "answer-first rate" ($$\Delta_{AFR}$$), which quantifies how often the gold answer is stated early in the chain. This metric is used both for mechanism analysis within a corpus (the "carrier experiment") and as a pre-training predictor across models. The prompt ablation study effectively disentangles the effect of answer visibility from the instruction to rationalize. The use of difference-in-differences (DiD) in the carrier experiment robustly controls for problem-subset effects. The methodology is well-designed to provide clear, actionable insights into a subtle but critical problem in LLM reasoning distillation.
The experimental evaluation is comprehensive and robust. The core finding of a significant accuracy drop (16.2 points on MATH-500, up to 27.2 points on AIME) due to answer-conditioned generation is demonstrated across multiple math benchmarks (GSM8K, Minerva, AIME) and even transfers to code generation (MBPP+, HumanEval+), where chains passing all tests still harm the student. Crucially, the effect is shown to be absent in multiple-choice knowledge tasks, aligning with the proposed mechanism that it affects derivation-based reasoning. The experiments involve eight thinking models from four distinct families (Qwen, DeepSeek, Llama-Nemotron, GLM), demonstrating the generality of the finding. The prediction of the penalty using the $$\Delta_{AFR}$$ signature before training, with a low mean absolute error of 2.9 points, is a powerful result. The length-matched controls effectively rule out chain length as the sole explanatory factor, reinforcing that the content and generation process are the culprits. Statistical significance is consistently reported using paired bootstrap and McNemar's test, adding to the credibility of the results.
The paper demonstrates a strong commitment to reproducibility. It explicitly states that code and data are released to regenerate every reported number. This includes exact generation prompts, matched problem identifiers, generated chains, metadata, full training configurations, and the evaluation harness. Correctness-critical components like the symbolic answer checker and answer-position matcher have pinned dependencies. A manifest maps every benchmark, condition, and seed cell to its evaluation record, ensuring traceability. The held-out penalty predictions were recorded *before* training, and the released code recomputes the cross-model fit, further enhancing trust. This level of detail and transparency is exemplary.
The paper acknowledges several limitations. The "one-bit experiment" primarily uses 935 matched math problems on one architecture family at a time, which, while controlled, limits the immediate generalization to all domains or problem types. The content-versus-length attribution is focused on MATH-500. The cross-model fit is based on a relatively small sample of eight models, and while robust, the Fisher-z interval on 'r' is wide. Some experiments, such as the code domain and student-output transfer, are not length-controlled, though other length-matched controls support the core claim. The single-seed rows are noted to support consistent sign rather than exact magnitude.
This paper has significant broader impact for the field of LLM development and application, particularly in self-improvement and reasoning distillation. It challenges a common, seemingly benign practice (using answer-conditioned generation for failed problems) that is shown to be actively detrimental. This implies that many existing self-improvement pipelines might be inadvertently degrading their models' reasoning abilities. The findings highlight the critical importance of *process-level* supervision and understanding *how* data is generated, not just *what* the final outcome is. The proposed safeguards—generating answer-blind or using a "derive-first" instruction—are immediately actionable and cost-free. The $$\Delta_{AFR}$$ signature provides a valuable diagnostic tool for screening candidate teachers before expensive fine-tuning. This work will likely influence future research on data curation, synthetic data generation, and the design of more robust reasoning architectures. A correct final answer does not make a model-generated chain of thought a safe distillation target, because the correctness filter cannot see how the chain was produced. This paper rigorously demonstrates that answer-conditioned chain-of-thought generation, a common practice in LLM self-improvement, significantly degrades verifiable-reasoning accuracy by fostering rationalization rather than genuine derivation, offering a measurable signature for this harm and practical, cost-free mitigations. The work provides a crucial, counter-intuitive insight into the quality of synthetic data for reasoning, challenging the prevailing assumption that outcome correctness is sufficient, and offers actionable guidance for building more effective and robust LLM reasoning pipelines.
Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilities: one-shot in-context imitation from human video demonstrations, on-the-fly policy improvement, robustness to perturbations, and stronger performance on multi-stage, long-horizon tasks. We also observe, for the first time, steady gains in closed-loop performance as pretraining context length scales. At its core, RoboTTT integrates Test-Time Training into robot foundation models such as Vision-Language-Action policies, yielding a sequence model whose recurrent state consists of fast weights, parameters updated by gradient descent during both training and inference, compressing histories into weight space and retrieving contextual information for long-context conditioning. To scale training context length, the recipe combines sequence action forcing with truncated backpropagation through time. On challenging real-robot manipulation tasks, RoboTTT improves overall performance by 87% over the single-step context baseline and fully completes a five-minute, ten-stage assembly task, which no baseline ever does. RoboTTT trained with 8K-timestep context outperforms the same model pretrained with 1K timesteps by 62%, suggesting context length as a new scaling axis for robot foundation models. Videos are available at https://research.nvidia.com/labs/gear/robottt/
Primary: NVIDIA
All Institutions: NVIDIA, Stanford Vision and Learning Lab
RoboTTT introduces a novel integration of Test-Time Training into robot foundation models, enabling 8K-timestep context scaling and unlocking new capabilities like one-shot in-context imitation and long-horizon task completion with significant performance gains.
The paper introduces RoboTTT, a significant architectural and training innovation for robot foundation models. By integrating Test-Time Training (TTT) into Vision-Language-Action (VLA) policies, the authors transform the recurrent state into "fast weights" that are updated via gradient descent during both training and inference. This allows the model to compress long histories (up to 8K timesteps) into its parameter space, effectively enabling in-context learning capabilities within a single policy instance. The training recipe combines sequence action forcing with truncated backpropagation through time (TBPTT) to handle the computational complexity of such long contexts. This approach is theoretically grounded in meta-learning principles and offers a compelling alternative to standard attention-based context windows or fixed-dimension recurrent states, particularly for maintaining low inference latency while scaling context.
The empirical evaluation is robust and demonstrates clear superiority over baselines. The paper reports an 87% improvement in overall performance on challenging real-robot manipulation tasks compared to single-step context baselines. A key highlight is the successful completion of a five-minute, ten-stage assembly task, a feat no baseline achieved. The scaling analysis is particularly strong, showing steady gains in closed-loop performance as pretraining context length increases, with an 8K-timestep model outperforming a 1K-timestep model by 62%. The ability to perform one-shot in-context imitation from human video demonstrations and on-the-fly policy improvement further validates the utility of the long-context conditioning. The results are presented on real-robot hardware, adding significant practical value over simulation-only studies.
The paper provides a detailed description of the training recipe, including the use of sequence action forcing and TBPTT, which are crucial for replicating the results. The availability of videos at the NVIDIA GEAR lab website enhances transparency regarding the robot's behavior. However, the code repository is not explicitly listed in the provided text (only the demo URL is given), which may hinder immediate reproducibility. The architectural details of integrating TTT into VLAs are described, but the specific hyperparameters for the fast weight updates and the exact structure of the VLA backbone would need to be verified against the full text or code release for perfect reproducibility.
The primary limitation is the computational cost associated with the TTT mechanism during inference, as it involves gradient updates. While the authors claim no growth in inference latency compared to baselines, this likely relies on specific hardware optimizations or approximations that may not generalize to all edge devices. Additionally, the performance gains are demonstrated on specific manipulation tasks; the generalizability to other robotic domains (e.g., locomotion, aerial robotics) or more complex, unstructured environments remains to be seen. The reliance on high-quality demonstration data for the in-context learning aspect could also be a bottleneck in data-scarce scenarios.
RoboTTT represents a step towards more autonomous and adaptable robot systems that can learn from experience in real-time. By enabling long-horizon planning and adaptation without retraining, it reduces the operational overhead of deploying robot policies. The ability to learn from human demonstrations via in-context learning could democratize robot programming, allowing non-experts to teach robots new skills simply by showing them. However, the increased complexity and computational demands may limit deployment to high-end hardware, potentially widening the gap in robotic capabilities between well-funded and resource-constrained settings. RoboTTT introduces a novel integration of Test-Time Training into robot foundation models, enabling 8K-timestep context scaling and unlocking new capabilities like one-shot in-context imitation and long-horizon task completion with significant performance gains.
Robotic manipulation is inherently multi-frame: local actions may be simple in an end-effector frame, while transport, upright-object handling, and whole-body coordination are better represented in a base-aligned frame. However, modern diffusion-based visuomotor policies typically commit to a single predefined action frame, forcing one denoiser to model action distributions that are often unnecessarily complex in that frame. We propose Mixture of Frames Policy (MoF), a diffusion policy that performs synchronized action denoising across multiple coordinate frames. MoF maintains a single canonical diffusion state, re-expresses it in several task-relevant frames, applies frame-specialized denoisers, and fuses their noise predictions back in the canonical frame. To make this possible for intermediate noisy diffusion states, we introduce a column-based 6D rotation representation within an SE(3) action parameterization that supports exact, differentiable frame transformations without requiring noisy rotations to lie on the SO(3) manifold. Across nine simulated bimanual manipulation tasks, we show that the best action frame is task-dependent and that MoF improves over oracle frame selection and standard Mixture-of-Experts (MoE) baselines. We further evaluate MoF on two real-world bimanual mobile manipulation tasks, demonstrating that it outperforms all constituent single-frame baselines. Project homepage: https://mofpo.github.io
Primary: Stanford University
All Institutions: Stanford University, Apple, Toyota Research Institute
The paper presents a highly innovative and technically sound approach to improving diffusion-based visuomotor policies by leveraging multi-frame reasoning. The introduction of a column-based 6D rotation representation for exact noisy state transformation is a significant methodological advance. Comprehensive experiments in simulation and on real-world robots demonstrate substantial improvements over strong baselines, including oracle frame selection. The work is well-motivated, rigorously evaluated, and addresses a fundamental limitation in current robotic policy architectures. It is a strong candidate for top-tier publication and will likely influence future research in robotic policy learning.
The paper proposes a novel architecture for diffusion-based visuomotor policies called Mixture of Frames (MoF). The core insight is that robotic manipulation actions are naturally represented in different coordinate frames at different stages of a task (e.g., end-effector for grasping, base for transport). Standard diffusion policies fix a single action frame, forcing the denoiser to model complex, frame-dependent distributions. MoF maintains a single canonical diffusion state but applies frame-specialized denoisers in their respective frames, fusing the noise predictions back into the canonical frame. A critical technical contribution is the introduction of a column-based 6D rotation representation for SE(3) actions. This allows for exact, differentiable frame transformations of *noisy* intermediate states without requiring projection onto the SO(3) manifold, which is non-linear and lossy. This mathematical insight is elegant and solves a significant implementation hurdle for multi-frame reasoning in diffusion policies. The method is generalizable to any diffusion policy and can be implemented as a drop-in replacement or extension.
The evaluation is comprehensive and rigorous. The authors first establish the motivation by showing that action frame choice has a substantial, task-dependent impact on performance in bimanual mobile manipulation, with a 15% gap between the best and worst single frames. They then demonstrate that MoF outperforms oracle frame selection (which knows the best frame per task) and standard Mixture-of-Experts (MoE) baselines that do not operate in multiple action frames. The ablation studies are particularly strong, showing that the proposed column-based representation is essential (orthogonalization leads to collapse in ensemble variants) and that the router effectively switches between frames based on task phase. Real-world experiments on two complex bimanual mobile manipulation tasks (Pouring, Serving) further validate the approach, showing significant improvements over single-frame baselines. The use of simulated benchmarks (BiGym, DexMimicGen) and real-world validation provides a robust evidence base.
The paper provides detailed descriptions of the method, including the specific rotation representation and the fusion mechanism. The project homepage link suggests code availability. The experimental setup is well-described, including dataset sizes, training epochs, and evaluation metrics. The ablation studies provide clear evidence of component contributions. The reliance on standard simulators (BiGym) and real-world hardware (HoMMI) enhances reproducibility.
The method relies on a designer-specified set of candidate frames. While the router learns to weight them, the set must be predefined. The paper acknowledges this and suggests future work on learning or discovering frames. Additionally, the router is supervised via denoising loss, which may not perfectly align with policy performance, though the results suggest it works well in practice. The current experiments focus on smaller diffusion policies; scaling to large VLA models is noted as future work.
This work has significant implications for robotic manipulation, particularly in complex, multi-stage tasks where different coordinate systems are advantageous. By improving the robustness and performance of visuomotor policies, it contributes to the broader goal of autonomous robots capable of operating in unstructured environments. The technical contribution regarding SE(3) transformations is also valuable for the broader machine learning community working with geometric deep learning and diffusion models. The paper presents a highly innovative and technically sound approach to improving diffusion-based visuomotor policies by leveraging multi-frame reasoning. The introduction of a column-based 6D rotation representation for exact noisy state transformation is a significant methodological advance. Comprehensive experiments in simulation and on real-world robots demonstrate substantial improvements over strong baselines, including oracle frame selection. The work is well-motivated, rigorously evaluated, and addresses a fundamental limitation in current robotic policy architectures. It is a strong candidate for top-tier publication and will likely influence future research in robotic policy learning.
Dexterous grasp generation across robot hands is challenging because hands differ in kinematic topology, actuation dimensions, and native command spaces. We introduce GraspGraphNet, a topology-aware grasp generation framework that represents each hand as a URDF-derived kinematic graph and directly generates executable palm poses and joint configurations. GraspGraphNet combines hierarchical object surface encoding, differentiable forward kinematics, and dynamic world-edge message passing to model evolving robot-object interactions. It applies conditional flow matching directly in executable palm-pose and joint-state space, avoiding post-processing optimization, inverse kinematics, and retargeting. Using a shared model trained on Barrett Hand, Allegro Hand, and Shadow Hand, GraspGraphNet achieves an average success rate of 83.48% with 40ms inference time per grasp on a 40-object benchmark. Without retraining, the same model achieves 72.70% success on controlled finger-removal variants, demonstrating robustness to hand-topology variations. These results suggest that graph-structured hand representations can effectively support dexterous grasp generation across robot hands with different kinematic structures. Project: https://lysees.github.io/graspgraphnet-page
Primary: Korea Advanced Institute of Science and Technology (KAIST)
All Institutions: Korea Advanced Institute of Science and Technology (KAIST)
GraspGraphNet presents a significant technical contribution to multi-embodiment robotic manipulation by introducing a topology-aware, graph-based flow matching framework that achieves state-of-the-art performance and efficiency in generating executable dexterous grasps across diverse robot hands.
The paper proposes GraspGraphNet, a novel framework for multi-embodiment dexterous grasp generation. The core innovation lies in representing robot hands as URDF-derived kinematic graphs and object surfaces as hierarchical point clouds. By integrating differentiable forward kinematics into a graph neural network (GNN) architecture, the model respects the topological constraints of each hand. Crucially, it employs dynamic world-edge message passing to update robot-object interaction features at every step of a conditional flow matching process. This allows the model to generate executable palm poses and joint configurations directly, bypassing the need for inverse kinematics or post-processing optimization typically required by contact-based methods. The use of flow matching for continuous state refinement is a sophisticated choice that addresses the multimodality of grasp distributions better than direct regression or standard diffusion models.
The experimental evaluation is comprehensive and rigorous. The authors evaluate on three distinct dexterous hands (Barrett, Allegro, Shadow) across 40 objects, providing a robust benchmark for multi-embodiment generalization. The results show state-of-the-art success rates (83.48% average) while significantly outperforming baselines in inference speed (40ms vs. 265ms+ for competitors). The finger-removal experiments are particularly compelling, demonstrating that the graph-based representation allows for zero-shot generalization to topologically modified hands (72.70% success), a feat previous methods struggle with. Real-world deployment on a Leap Hand further validates the practical utility of the approach. The ablation studies effectively isolate the contributions of dynamic world edges, cross-attention, and the flow-matching objective.
The paper provides sufficient implementation details, including network architecture dimensions, training hyperparameters (AdamW, learning rate, epochs), and dataset sources (CMapDataset). The use of standard simulation environments (Isaac Gym) and publicly available datasets enhances reproducibility. The inclusion of a project page suggests code availability, which is critical for replication in this domain.
The primary limitation is the reliance on URDF definitions, which requires accurate kinematic models for each hand. While this is standard in simulation, real-world deployment may require precise calibration or adaptation to handle kinematic drift or non-rigid deformations not captured in the URDF. Additionally, the model is evaluated on a specific set of hands; generalization to radically different morphologies (e.g., soft robotic hands or grippers with vastly different joint limits) is not demonstrated, though the authors acknowledge this as future work. The reliance on flow matching also introduces a small number of integration steps (K=3), which, while efficient, may occasionally produce less optimal grasps compared to methods with more extensive search or optimization phases, although the speed trade-off is favorable.
This work significantly advances the field of robotic manipulation by providing a unified, efficient, and robust method for dexterous grasp generation across diverse hardware. By eliminating the need for hand-specific retargeting and optimization, it lowers the barrier to deploying dexterous manipulation policies on new robotic platforms. The graph-based approach offers a principled way to handle structural variations, which could inspire similar topology-aware representations in other robotic control tasks. The efficiency gains (40ms inference) make real-time dexterous manipulation more feasible for practical applications. GraspGraphNet presents a significant technical contribution to multi-embodiment robotic manipulation by introducing a topology-aware, graph-based flow matching framework that achieves state-of-the-art performance and efficiency in generating executable dexterous grasps across diverse robot hands.
Bipedal robots are challenging to control because they operate close to instability, where small variations in foot-terrain contact can rapidly destabilize locomotion. On rigid terrain, bipedal robots mitigate this fragility by using well-established contact mechanics and control strategies. On flowable surfaces such as granular slopes, foot contact can induce large surface deformations and solid-fluid-like transitions, coupling terrain effects with robot dynamics, leading to underperformance or failure. This is partly due to the lack of reliable methods to represent the dynamics of flowable terrain, making it difficult to account for terrain effects in locomotion design. Here, we investigate how controlling terrain response can improve bipedal locomotion on granular slopes by studying the terradynamics of cleated feet, thin plates emanating from the foot soles. Systematic studies of a small-scale (1.4 kg) robophysical biped reveal that cleats with sparse and dense spacing lead to excessive terrain yielding and resistance, respectively, degrading performance and leading to failure. An intermediate cleat spacing distributes interaction forces to maintain substrate stresses near (or below) the yield threshold, enabling walking on granular slopes up to 30 degrees. Guided by these principles, we design a foot that actively adjusts cleat depth and accommodates both rigid and granular terrain. We also demonstrate that the principles of effective foot-terrain interaction translate to a larger (15 kg) autonomous biped. Our study presents an alternative to conventional body-centric robot control approaches, which regulate terrain-induced effects through body motion, by instead regulating terrain interactions through limb-centric approach.
Primary: Georgia Institute of Technology
All Institutions: Georgia Institute of Technology, Northeastern University
This paper presents a seminal contribution to robotic locomotion on deformable terrain, establishing a mechanistic framework for cleated-foot interaction that enables robust bipedal walking on steep granular slopes. By systematically characterizing terradynamics and validating findings across multiple scales, the authors provide a foundational understanding that bridges the gap between granular physics and legged robot design, offering a pathway to more capable autonomous systems in challenging natural environments.
The paper employs a rigorous robophysical approach, combining systematic experimental variation of foot morphology (cleat spacing and depth) with granular terradynamics analysis. The methodology is grounded in physical principles rather than purely data-driven learning, utilizing Particle Image Velocimetry (PIV) and dual-plate intrusion tests to elucidate the underlying mechanics of granular flow. The introduction of a limb-centric control paradigm, where terrain interaction is regulated through foot design rather than just body dynamics, represents a significant conceptual shift in legged locomotion research. The integration of these insights into an adaptive foot mechanism (retractable cleats) and validation on a large-scale autonomous biped (HECTOR) demonstrates a robust engineering methodology.
The experimental evaluation is comprehensive and multi-scaled. It begins with a small-scale constrained robot (BLUEY) to isolate variables, proceeds to granular intrusion tests for mechanistic understanding, and culminates in challenging field-relevant scenarios with a large-scale autonomous robot (HECTOR). The use of PIV to visualize particle flow provides unique, high-quality data linking foot geometry to substrate response. The results clearly demonstrate that intermediate cleat spacing optimizes performance by balancing traction and substrate disturbance, a finding that is both counter-intuitive and practically significant. The successful translation of these principles to a 15 kg robot is a strong validation of the findings.
The paper provides detailed descriptions of the robot kinematics, gait generation (ZMP-based), and experimental setups. The inclusion of pseudocode for the adaptive cleat control and specific parameters for the granular substrate (poppy seeds, volume fraction) enhances reproducibility. However, the reliance on specific robophysical platforms and the complexity of the granular terrain preparation (air-fluidization) may pose challenges for exact replication without access to similar infrastructure. The open-source nature of the HECTOR platform aids in this regard.
The study is limited to rectangular cleats and specific granular media (poppy seeds). The scaling laws derived may not directly apply to human-scale robots or different substrate types (e.g., cohesive soils) without further validation. The gait strategies tested are relatively simple (quasi-static for BLUEY, limited MPC exploration for HECTOR), leaving room for optimization. The authors acknowledge these limitations and suggest future work involving DEM simulations and more complex terrain.
This work has significant implications for the field of legged robotics, particularly for applications in unstructured environments such as search and rescue, planetary exploration, and agricultural robotics. By providing a mechanistic framework for locomotion on flowable terrain, it enables the design of more robust and versatile robotic systems. The shift from body-centric to limb-centric control strategies offers a new paradigm for handling terrain-robot coupling, potentially influencing future research in adaptive morphology and control. This paper presents a seminal contribution to robotic locomotion on deformable terrain, establishing a mechanistic framework for cleated-foot interaction that enables robust bipedal walking on steep granular slopes. By systematically characterizing terradynamics and validating findings across multiple scales, the authors provide a foundational understanding that bridges the gap between granular physics and legged robot design, offering a pathway to more capable autonomous systems in challenging natural environments.
Video diffusion transformers (vDiTs) generate high quality video but introduce extremely high compute cost due to the long diffusion timesteps and self attention computation. As diffusion timesteps are reduced, the computation cost of self attention becomes the dominant bottleneck. Existing acceleration approaches largely inherit sparse attention techniques from large language models, which fail to consider the unique spatiotemporal correlation of video data. This paper presents Kaleido, an algorithm hardware codesign that accelerates all operations in vDiTs by exploiting channel-wise spatiotemporal correlations in latent space. Based on this insight, we propose a lightweight channelwise reuse algorithm that skips redundant computations by reusing partial results while preserving higher generative quality than prior methods (>17 dB). To efficiently support this algorithm, we design a systolic array like accelerator with reconfigurable processing elements and a lightweight data dispatcher to mitigate irregular sparsity and data access patterns introduced by our reuse algorithm. Evaluations across three mainstream vDiT models show that Kaleido achieves up to 5.9x speedup and 16.0x energy savings over state of the art accelerators.
Primary: Unknown
All Institutions: Unknown
Kaleido presents a compelling algorithm-hardware co-design that significantly accelerates Video Diffusion Transformers by exploiting latent space correlations, offering substantial speedups and energy savings while maintaining high generative quality, thereby addressing a critical bottleneck in the deployment of advanced video generation models.
The paper proposes "Kaleido," an algorithm-hardware co-design specifically targeting Video Diffusion Transformers (vDiTs). The core algorithmic insight is the exploitation of channel-wise spatiotemporal correlations in the latent space to enable computation skipping (reusing partial results). This is a significant methodological shift from standard sparse attention techniques used in LLMs, which do not account for the unique redundancy in video generation tasks. The hardware design features a systolic array-like accelerator with reconfigurable processing elements and a lightweight data dispatcher to handle the irregular sparsity introduced by the algorithm. The approach is technically sound, addressing the dominant bottleneck of self-attention in long-timestep diffusion processes. However, the novelty is somewhat incremental in the broader context of hardware acceleration, as algorithm-hardware co-design for transformers is a well-trodden path, though the specific application to vDiTs with this correlation-based reuse is a strong contribution.
The evaluation demonstrates up to 5.9x speedup and 16.0x energy savings compared to state-of-the-art accelerators across three mainstream vDiT models. The claim of preserving higher generative quality (>17 dB improvement in PSNR/SSIM relative to prior skipping methods) is a strong empirical result. The use of standard metrics and comparison against SOTA baselines provides credibility. However, the lack of specific details on the baseline models and the exact nature of the "state of the art" accelerators in the abstract limits the immediate assessability of the magnitude of the improvement. The results are promising and suggest significant practical utility for deploying vDiTs.
The paper is published on arXiv, and while the abstract provides a high-level overview, the full text would need to be scrutinized for hyperparameters, model configurations, and hardware specifications to ensure reproducibility. The mention of "lightweight" components suggests a focus on practical implementation, which often aids reproducibility if code is released. Without code availability explicitly stated in the abstract, reproducibility is assumed to be moderate to high based on the detailed algorithmic description implied.
The primary limitation is the reliance on the assumption that channel-wise spatiotemporal correlations are consistent across different vDiT architectures and video domains. The performance gain might vary depending on the specific video content and the complexity of the diffusion process. Additionally, the hardware design's flexibility (reconfigurable PEs) might come with overheads that are not fully detailed in the abstract. The comparison is against "state of the art accelerators," which may include specialized custom hardware that is not widely available, potentially skewing the perceived advantage.
This work has significant implications for the democratization of high-quality video generation by reducing the computational and energy barriers. It enables faster iteration for researchers and more accessible deployment for developers. The focus on energy efficiency also aligns with the growing need for sustainable AI. By addressing the specific bottlenecks of vDiTs, it paves the way for more complex and longer video generation tasks. Kaleido presents a compelling algorithm-hardware co-design that significantly accelerates Video Diffusion Transformers by exploiting latent space correlations, offering substantial speedups and energy savings while maintaining high generative quality, thereby addressing a critical bottleneck in the deployment of advanced video generation models.
Modern edge system-on-chips (SoCs) combine CPUs, integrated GPUs (iGPUs), and neural processing units (NPUs), yet existing LLM runtimes typically make coarse device-level decisions or optimize operators in isolation. As a result, they underutilize heterogeneous resources, particularly on unified-memory platforms where performance depends on both device placement and task-graph coordination. We present HeteroMosaic, a heterogeneity-first scheduling framework for edge LLM inference. HeteroMosaic first uses a heterogeneous roofline model to identify when combining iGPU and NPU execution is beneficial. It then decomposes inference into dependency-preserving micro-batches that expose cross-accelerator overlap and applies trace-guided co-optimization of scheduling and device allocation under practical effects such as memory contention, DVFS, device variation, and NPU runtime overheads. We implement HeteroMosaic in PyTorch C++ and evaluate it on three AMD Ryzen AI platforms spanning NPU-heavy, balanced, and iGPU-heavy designs. On the balanced platform, HeteroMosaic achieves up to 1.73X speedup over an iGPU baseline, 1.78X over an NPU baseline, and 2.05X over frameworks such as llama.cpp, while reducing energy by up to 45.3%. It also improves performance over prior heterogeneous edge AI solutions by up to 2.35X.
Primary: University of Illinois Urbana-Champaign (UIUC)
All Institutions: University of Illinois Urbana-Champaign (UIUC), Advanced Micro Devices, Inc. (AMD)
HeteroMosaic introduces a novel scheduling framework for edge LLM inference that significantly outperforms existing baselines by exploiting fine-grained heterogeneous execution opportunities on unified-memory SoCs, offering a practical path toward energy-efficient, high-performance edge AI deployment.
The paper proposes HeteroMosaic, a scheduling framework that treats edge LLM inference as a heterogeneity-first problem rather than a simple placement problem. The core methodological contributions include: 1) A heterogeneous roofline model to theoretically bound the benefits of combining iGPU and NPU execution; 2) Dependency-preserving micro-batching to expose cross-accelerator overlap opportunities that coarse-grained batching misses; 3) Trace-guided co-optimization of scheduling and device allocation that accounts for real-world system effects like memory contention, DVFS, and NPU runtime overheads. This approach is technically sophisticated, moving beyond static operator mapping to dynamic, graph-level coordination. The insight that fine-grained heterogeneity can be exploited via scheduling on unified-memory platforms is a significant shift in how edge AI systems are designed.
The evaluation is conducted on three distinct AMD Ryzen AI platforms (NPU-heavy, balanced, iGPU-heavy), providing strong empirical evidence for the framework's adaptability. The results show substantial improvements: up to 1.73x speedup over iGPU-only, 1.78x over NPU-only, and 2.05x over llama.cpp on a balanced platform, with up to 45.3% energy reduction. The comparison against strong baselines (including existing heterogeneous solutions) demonstrates clear superiority. The use of real hardware rather than simulation adds significant credibility. The evaluation covers multiple models, ensuring the findings are not model-specific artifacts.
The authors implement HeteroMosaic in PyTorch C++, which is a standard and accessible framework for such systems research. The paper details the architectural properties required (programmable accelerators, shared memory, low-overhead sync) and the specific optimizations applied. While the exact proprietary NPU driver details might be AMD-specific, the scheduling abstraction and the trace-guided optimization methodology are described with sufficient detail for replication on similar unified-memory SoCs. The codebase structure suggests it could be open-sourced or made available upon request, which is common for top-tier systems venues.
The current implementation is tightly coupled with AMD Ryzen AI platforms. The "heterogeneous roofline model" and scheduling heuristics may need significant re-tuning for other architectures (e.g., Qualcomm Hexagon, Apple Neural Engine) due to differences in NPU capabilities, memory bandwidth, and runtime overheads. The paper acknowledges that extending this to datacenter systems would require explicit communication modeling for PCIe/CXL/NVLink, which is a non-trivial extension not covered here. Additionally, the focus is on inference; training on such heterogeneous edge devices remains an open challenge.
This work has profound implications for the deployment of Large Language Models on edge devices, enabling faster and more energy-efficient AI on smartphones, laptops, and IoT devices. By demonstrating that fine-grained heterogeneity can be practically exploited, it encourages hardware designers to prioritize balanced accelerator designs and efficient interconnects. It also provides a blueprint for runtime systems to better utilize the increasingly heterogeneous compute resources available in modern SoCs, potentially reducing the carbon footprint of AI inference. HeteroMosaic introduces a novel scheduling framework for edge LLM inference that significantly outperforms existing baselines by exploiting fine-grained heterogeneous execution opportunities on unified-memory SoCs, offering a practical path toward energy-efficient, high-performance edge AI deployment.